Decoding method, apparatus and recording medium

0Associated
Cases 
0Associated
Defendants 
0Accused
Products 
0Forward
Citations 
0
Petitions 
0
Assignments
First Claim
1. A decoding method, implemented by a decoding apparatus having processing circuitry, comprising:
 where p is an integer equal to or greater than 1,decoding, by the processing circuitry, input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];
with a frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] being the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p], executing, by the processing circuitry, a parameter sequence conversion step of determining a converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] using the frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] as input to thereby generate the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] as a decoded approximate LSP parameter sequence {circumflex over (
)}θ
_{app}[1], {circumflex over (
)}θ
_{app}[2], . . . , {circumflex over (
)}θ
_{app}[p];
generating, by the processing circuitry, a decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{γ}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p] by converting the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p] into linear prediction coefficients;
calculating, by the processing circuitry, a decoded smoothed power spectral envelope series {circumflex over (
)}W_{65}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N] which is a series in frequency domain corresponding to the decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{γ}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p];
generating, by the processing circuitry, decoded sound signals using a frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N];
decoding, by the processing circuitry, input LSP codes to obtain a decoded LSP parameter sequence {circumflex over (
)}θ
[1], {circumflex over (
)}θ
[2], . . . , {circumflex over (
)}θ
[p]; and
decoding, by the processing circuitry, input time domain signal codes, and generating decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence for a preceding time segment or the decoded approximate LSP parameter sequence for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment,whereinthe processing circuitry determines a value of each converted frequency domain parameter ˜
ω
[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] through linear transformation which is based on a relationship of values between ω
[i] and one or more frequency domain parameters adjacent to ω
[i].
0 Assignments
0 Petitions
Accused Products
Abstract
The present invention reduces encoding distortion in frequency domain encoding compared to conventional techniques, and obtains LSP parameters that correspond to quantized LSP parameters for the preceding frame and are to be used in time domain encoding from coefficients equivalent to linear prediction coefficients resulting from frequency domain encoding. When p is an integer equal to or greater than 1, a linear prediction coefficient sequence which is obtained by linear prediction analysis of audio signals in a predetermined time segment is represented as a[1], a[2], . . . , a[p], and ω[1], ω[2], . . . , ω[p] are a frequency domain parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], an LSP linear transformation unit (300) determines the value of each converted frequency domain parameter ˜ω[i] (i=1, 2, . . . , p) in a converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p] using the frequency domain parameter sequence ω[1], ω[2], . . . , ω[p] as input, through linear transformation which is based on the relationship of values between ω[i] and one or more frequency domain parameters adjacent to ω[i].
38 Citations
No References
AUDIO ENCODER AND DECODER  
Patent #
US 20100286990A1
Filed 12/30/2008

Current Assignee
Dolby International AB

Sponsoring Entity
Dolby International AB

VOICE BAND EXPANSION DEVICE, VOICE BAND EXPANSION METHOD, AND COMMUNICATION APPARATUS  
Patent #
US 20100318350A1
Filed 05/11/2010

Current Assignee
Fujitsu Limited

Sponsoring Entity
Fujitsu Limited

Scalable and embedded codec for speech and audio signals  
Patent #
US 20080052068A1
Filed 08/10/2007

Current Assignee
David Campana, Robert Zopf, Robert Dunn, Robert Mcaulay, Joseph Aguilar, Craig Watkins, Chen JuinHwey, Xiaoquin Sun, Wang Wei

Sponsoring Entity
David Campana, Robert Zopf, Robert Dunn, Robert Mcaulay, Joseph Aguilar, Craig Watkins, Chen JuinHwey, Xiaoquin Sun, Wang Wei

Speech processing apparatus and mobile communication terminal  
Patent #
US 7,330,813 B2
Filed 08/05/2003

Current Assignee
Fujitsu Connected Technologies Limited

Sponsoring Entity
Fujitsu Limited

Timewarping frames of wideband vocoder  
Patent #
US 20080052065A1
Filed 08/22/2006

Current Assignee
Qualcomm Inc.

Sponsoring Entity
Qualcomm Inc.

Scalable and embedded codec for speech and audio signals  
Patent #
US 7,272,556 B1
Filed 09/23/1998

Current Assignee
AlcatelLucent SA

Sponsoring Entity
Lucent Technologies Inc.

Speech encoding at variable bit rate  
Patent #
US 5,933,803 A
Filed 12/05/1997

Current Assignee
Nokia Mobile Phones UK Limited

Sponsoring Entity
Nokia Mobile Phones UK Limited

Speech Processing apparatus and mobile communication terminal  
Patent #
US 20040042622A1
Filed 08/05/2003

Current Assignee
Fujitsu Connected Technologies Limited

Sponsoring Entity
Fujitsu Limited

Speech synthesis with equal interval line spectral pair frequency interpolation  
Patent #
US 5,864,796 A
Filed 02/06/1997

Current Assignee
Sony Corporation

Sponsoring Entity
Sony Corporation

Coding of a speech or music signal with quantization of harmonics components specifically and then residue components  
Patent #
US 5,806,024 A
Filed 12/23/1996

Current Assignee
NEC Corporation

Sponsoring Entity
NEC Corporation

Filter for speech modification or enhancement, and various apparatus, systems and method using same  
Patent #
US 5,822,732 A
Filed 05/02/1996

Current Assignee
Mitsubishi Electric Corporation

Sponsoring Entity
Mitsubishi Electric Corporation

Speech approximation using successive sinusoidal overlapadd models and pitchscale modifications  
Patent #
US 5,504,833 A
Filed 05/04/1994

Current Assignee
Georgia Tech Research Corporation

Sponsoring Entity
Georgia Tech Research Corporation

Audio analysis/synthesis system  
Patent #
US 5,327,518 A
Filed 08/22/1991

Current Assignee
Georgia Tech Research Corporation

Sponsoring Entity
Georgia Tech Research Corporation

Voice coding apparatus  
Patent #
US 5,003,604 A
Filed 03/09/1989

Current Assignee
Fujitsu Limited

Sponsoring Entity
Fujitsu Limited

Timewarping frames of wideband vocoder  
Patent #
US 8,239,190 B2
Filed 08/22/2006

Current Assignee
Qualcomm Inc.

Sponsoring Entity
Qualcomm Inc.

Voice band expansion device, voice band expansion method, and communication apparatus  
Patent #
US 8,280,727 B2
Filed 05/11/2010

Current Assignee
Fujitsu Limited

Sponsoring Entity
Fujitsu Limited

Audio encoder and decoder with long term prediction  
Patent #
US 8,494,863 B2
Filed 12/30/2008

Current Assignee
Dolby International AB

Sponsoring Entity
Dolby Laboratories Incorporated

ENCODING METHOD, ENCODER, PERIODIC FEATURE AMOUNT DETERMINATION METHOD, PERIODIC FEATURE AMOUNT DETERMINATION APPARATUS, PROGRAM AND RECORDING MEDIUM  
Patent #
US 20130311192A1
Filed 01/18/2012

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

ENCODING METHOD, DECODING METHOD, ENCODER, DECODER, PROGRAM, AND RECORDING MEDIUM  
Patent #
US 20130317814A1
Filed 02/08/2012

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

Packet Loss Concealment for Speech Coding  
Patent #
US 20140156267A1
Filed 02/07/2014

Current Assignee
Huawei Technologies Co. Ltd.

Sponsoring Entity
Huawei Technologies Co. Ltd.

Methods and Systems for Applications for Znumbers  
Patent #
US 20140201126A1
Filed 03/18/2014

Current Assignee
Z Advanced Computing Inc.

Sponsoring Entity
Z Advanced Computing Inc.

Scalable and embedded codec for speech and audio signals  
Patent #
US 9,047,865 B2
Filed 08/10/2007

Current Assignee
AlcatelLucent SA

Sponsoring Entity
AlcatelLucent SA

ENCODING METHOD, ENCODER, PROGRAM AND RECORDING MEDIUM  
Patent #
US 20150187366A1
Filed 09/30/2013

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

Scalable And Embedded Codec For Speech And Audio Signals  
Patent #
US 20150302859A1
Filed 05/04/2015

Current Assignee
AlcatelLucent SA

Sponsoring Entity
AlcatelLucent SA

Encoding method for acquiring codes corresponding to prediction residuals, decoding method for decoding codes corresponding to noise or pulse sequence, encoder, decoder, program, and recording medium  
Patent #
US 9,230,554 B2
Filed 02/08/2012

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

Packet loss concealment for speech coding  
Patent #
US 9,336,790 B2
Filed 02/07/2014

Current Assignee
Huawei Technologies Co. Ltd.

Sponsoring Entity
Huawei Technologies Co. Ltd.

CONTEXTBASED DATA CLASSIFICATION  
Patent #
US 20160292445A1
Filed 03/18/2016

Current Assignee
SECUDE AG

Sponsoring Entity
SECUDE AG

SYSTEM AND METHOD FOR INFRASONIC CARDIAC MONITORING  
Patent #
US 20160361041A1
Filed 06/14/2016

Current Assignee
The Research Foundation for The State University of New York

Sponsoring Entity
The Research Foundation for The State University of New York

Encoding method, encoder, program and recording medium  
Patent #
US 9,524,725 B2
Filed 09/30/2013

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

CONTEXTSENSITIVE COPY AND PASTE BLOCK  
Patent #
US 20170154188A1
Filed 01/17/2017

Current Assignee
SECUDE AG

Sponsoring Entity
Rainer Lindemann, Philipp Meier

System and method for updating an adaptive speech recognition model  
Patent #
US 9,697,822 B1
Filed 04/28/2014

Current Assignee
Apple Inc.

Sponsoring Entity
Apple Inc.

Encoding method, encoder, periodic feature amount determination method, periodic feature amount determination apparatus, program and recording medium  
Patent #
US 9,711,158 B2
Filed 01/18/2012

Current Assignee
Nippon Telegraph and Telephone Corporation

Sponsoring Entity
Nippon Telegraph and Telephone Corporation

FREQUENCY DOMAIN PARAMETER SEQUENCE GENERATING METHOD, ENCODING METHOD, DECODING METHOD, FREQUENCY DOMAIN PARAMETER SEQUENCE GENERATING APPARATUS, ENCODING APPARATUS, DECODING APPARATUS, PROGRAM, AND RECORDING MEDIUM  
Patent #
US 20170249947A1
Filed 02/16/2015

Current Assignee
Nippon Telegraph and Telephone Corporation, University of Tokyo

Sponsoring Entity
Nippon Telegraph and Telephone Corporation, University of Tokyo

APPROXIMATE VALUE ITERATION WITH COMPLEX RETURNS BY BOUNDING  
Patent #
US 20180012137A1
Filed 11/22/2016

Current Assignee
The Research Foundation for The State University of New York

Sponsoring Entity
The Research Foundation for The State University of New York

Method and system for feature detection  
Patent #
US 9,916,538 B2
Filed 03/18/2014

Current Assignee
Z Advanced Computing Inc.

Sponsoring Entity
Z Advanced Computing Inc.

SEMISUPERVISED AUTOENCODER FOR SENTIMENT ANALYSIS  
Patent #
US 20180165554A1
Filed 12/11/2017

Current Assignee
The Research Foundation for The State University of New York

Sponsoring Entity
The Research Foundation for The State University of New York

Frequency domain parameter sequence generating method, encoding method, decoding method, frequency domain parameter sequence generating apparatus, encoding apparatus, decoding apparatus, program, and recording medium  
Patent #
US 10,332,533 B2
Filed 02/16/2015

Current Assignee
Nippon Telegraph and Telephone Corporation, University of Tokyo

Sponsoring Entity
Nippon Telegraph and Telephone Corporation, University of Tokyo

FRAMEWORK AND METHODS OF DIVERSE EXPLORATION FOR FAST AND SAFE POLICY IMPROVEMENT  
Patent #
US 20190228309A1
Filed 01/24/2019

Current Assignee
The Research Foundation for The State University of New York

Sponsoring Entity
The Research Foundation for The State University of New York

5 Claims
 1. A decoding method, implemented by a decoding apparatus having processing circuitry, comprising:
where p is an integer equal to or greater than 1, decoding, by the processing circuitry, input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];with a frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] being the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p], executing, by the processing circuitry, a parameter sequence conversion step of determining a converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] using the frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] as input to thereby generate the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] as a decoded approximate LSP parameter sequence {circumflex over (
)}θ
_{app}[1], {circumflex over (
)}θ
_{app}[2], . . . , {circumflex over (
)}θ
_{app}[p];generating, by the processing circuitry, a decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{γ}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p] by converting the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p] into linear prediction coefficients;calculating, by the processing circuitry, a decoded smoothed power spectral envelope series {circumflex over (
)}W_{65}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N] which is a series in frequency domain corresponding to the decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{γ}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p];generating, by the processing circuitry, decoded sound signals using a frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N];decoding, by the processing circuitry, input LSP codes to obtain a decoded LSP parameter sequence {circumflex over (
)}θ
[1], {circumflex over (
)}θ
[2], . . . , {circumflex over (
)}θ
[p]; anddecoding, by the processing circuitry, input time domain signal codes, and generating decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence for a preceding time segment or the decoded approximate LSP parameter sequence for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment, wherein the processing circuitry determines a value of each converted frequency domain parameter ˜
ω
[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] through linear transformation which is based on a relationship of values between ω
[i] and one or more frequency domain parameters adjacent to ω
[i]. View Dependent Claims (5)
 2. A decoding method, implemented by a decoding apparatus having processing circuitry, comprising:
where p is an integer equal to or greater than 1, decoding, by the processing circuitry, input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];with a frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] being the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p], executing, by the processing circuitry, a parameter sequence conversion step of determining a converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] using the frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] as input to thereby generate the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] as a decoded approximate LSP parameter sequence {circumflex over (
)}θ
_{app}[1], {circumflex over (
)}θ
_{app}[2], . . . , {circumflex over (
)}θ
_{app}[p];calculating, by the processing circuitry, a decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N] based on the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];generating, by the processing circuitry, decoded sound signals using a frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N];decoding, by the processing circuitry, input LSP codes to obtain a decoded LSP parameter sequence {circumflex over (
)}θ
[1], {circumflex over (
)}θ
[2], . . . , {circumflex over (
)}θ
[p]; anddecoding, by the processing circuitry, input time domain signal codes, and generating decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence for a preceding time segment or the decoded approximate LSP parameter sequence for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment, wherein the processing circuitry determines a value of each converted frequency domain parameter ˜
ω
[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] through linear transformation which is based on a relationship of values between ω
[i] and one or more frequency domain parameters adjacent to ω
[i].
 3. A decoding apparatus comprising:
where p is an integer equal to or greater than 1, an adjusted LSP code decoding unit that decodes input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];a decoded LSP linear transformation unit that, with a frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] being the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p], executes a parameter sequence converting unit of determining a converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] using the frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] as input to thereby generate the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] as a decoded approximate LSP parameter sequence {circumflex over (
)}θ
_{app}[1], {circumflex over (
)}θ
_{app}[2], . . . , {circumflex over (
)}θ
_{app}[p];a decoded linear prediction coefficient sequence generating unit that generates a decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{γ}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p] by converting the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p] into linear prediction coefficients;a decoded smoothed power spectral envelope series calculating unit that calculates a decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N] which is a series in frequency domain corresponding to the decoded adjusted linear prediction coefficient sequence {circumflex over (
)}a_{65}[1], {circumflex over (
)}a_{γ}[2], . . . , {circumflex over (
)}a_{γ}[p];a frequency domain decoding unit that generates decoded sound signals using a frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N];an LSP code decoding unit that decodes input LSP codes to obtain a decoded LSP parameter sequence {circumflex over (
)}θ
[1], {circumflex over (
)}θ
[2], . . . , {circumflex over (
)}θ
[p]; anda time domain decoding unit that decodes input time domain signal codes, and generates decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence obtained by the LSP code decoding unit for a preceding time segment or the decoded approximate LSP parameter sequence obtained in the decoded LSP linear transformation unit for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment, wherein the parameter sequence conversion unit determines a value of each converted frequency domain parameter ˜
ω
[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] through linear transformation which is based on a relationship of values between ω
[i] and one or more frequency domain parameters adjacent to ω
[i].
 4. A decoding apparatus comprising:
where p is an integer equal to or greater than 1, an adjusted LSP code decoding unit that decodes input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];a decoded LSP linear transformation unit that, with a frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] being the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}_{γ}[p], executes a parameter sequence converting unit of determining a converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] using the frequency domain parameter sequence ω
[1], ω
[2], . . . , ω
[p] as input to thereby generate the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] as a decoded approximate LSP parameter sequence {circumflex over (
)}θ
_{app}[1], {circumflex over (
)}θ
_{app}[2], . . . , {circumflex over (
)}θ
_{app}[p];a decoded smoothed power spectral envelope series calculating unit that calculates a decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N] based on the decoded adjusted LSP parameter sequence {circumflex over (
)}θ
_{γ}[1], {circumflex over (
)}θ
_{γ}[2], . . . , {circumflex over (
)}θ
_{γ}[p];a frequency domain decoding unit that generates decoded sound signals using a frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over (
)}W_{γ}[1], {circumflex over (
)}W_{γ}[2], . . . , {circumflex over (
)}W_{γ}[N];an LSP code decoding unit that decodes input LSP codes to obtain a decoded LSP parameter sequence {circumflex over (
)}θ
[1], {circumflex over (
)}θ
[2], . . . , {circumflex over (
)}θ
[p]; andan time domain decoding unit that decodes input time domain signal codes, and generates decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence obtained in the LSP code decoding unit for a preceding time segment or the decoded approximate LSP parameter sequence obtained in the decoded LSP linear transformation unit for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment, wherein the parameter sequence conversion unit determines a value of each converted frequency domain parameter ˜
ω
[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜
ω
[1], ˜
ω
[2], . . . , ˜
ω
[p] through linear transformation which is based on a relationship of values between ω
[i] and one or more frequency domain parameters adjacent to ω
[i].
1 Specification
This application is a continuation of and claims the benefit of priority under 35 U.S.C. § 120 from U.S. application Ser. No. 16/398,429 filed Apr. 30, 2019, which is a continuation of U.S. application Ser. No. 15/302,094 filed May 16, 2017 (now U.S. Pat. No. 10,332,533 issued Jun. 25, 2019), the entire contents of which are incorporated herein by reference. U.S. application Ser. No. 15/302,094 is a National Stage of PCT/JP2015/054135 filed Feb. 16, 2015, which claims the benefit of priority under 35 U.S.C. § 119 from Japanese Application No. 2014089895 filed Apr. 24, 2014.
The present invention relates to encoding techniques, and more particularly to techniques for converting frequency domain parameters equivalent to linear prediction coefficients.
In encoding of speech or sound signals, schemes that perform encoding using linear prediction coefficients obtained by linear prediction analysis of input sound signals are widely employed.
For instance, according to NonPatent Literatures 1 and 2, input sound signals in each frame are coded by either a frequency domain encoding method or a time domain encoding method. Whether to use the frequency domain or time domain encoding method is determined in accordance with the characteristics of the input sound signals in each frame.
Both in the time domain and frequency domain encoding methods, linear prediction coefficients obtained by linear prediction analysis of input sound signal are converted to a sequence of LSP parameters, which is then coded to obtained LSP codes, and also a quantized LSP parameter sequence corresponding to the LSP codes is generated. In the time domain encoding method, encoding is carried out by using linear prediction coefficients determined from a quantized LSP parameter sequence for the current frame and a quantized LSP parameter sequence for the preceding frame as the filter coefficients for a synthesis filter serving as a timedomain filter, applying the synthesis filter to a signal generated by synthesis of the waveforms contained in an adaptive codebook and the waveforms contained in a fixed codebook so as to determine a synthesized signal, and determining indices for the respective codebooks such that the distortion between the synthesized signal determined and the input sound signal is minimized.
In the frequency domain encoding method, a quantized LSP parameter sequence is converted to linear prediction coefficients to determine a quantized linear prediction coefficient sequence; the quantized linear prediction coefficient sequence is smoothed to determine a adjusted quantized linear prediction coefficient sequence; a signal from which the effect of the spectral envelope has been removed is determined by normalizing each value in a frequency domain signal series which is determined by converting the input sound signal to the frequency domain using each value in a power spectral envelope series, which is a series in the frequency domain corresponding to the adjusted quantized linear prediction coefficients; and the determined signal is coded by variable length encoding taking into account spectral envelope information.
As described, linear prediction coefficients determined through linear prediction analysis of the input sound signal are employed in common in the frequency domain and time domain encoding methods. Linear prediction coefficients are converted into a sequence of frequency domain parameters equivalent to the linear prediction coefficients, such as LSP (Line Spectrum Pair) parameters or ISP (Immittance Spectrum Pairs) parameters. Then, LSP codes (or ISP codes) generated by encoding the LSP parameter sequence (or ISP parameter sequence) are transmitted to a decoding apparatus. The frequencies from 0 to π of LSP parameters used in quantization or interpolation are sometimes specifically referred distinctively as LSP frequencies (LSF) or as ISP frequencies (ISF) in the case of ISP frequencies; however, such frequency parameters are referred to as LSP parameters or ISP parameters in the description of the present application.
Referring to
In the following description, an LSP parameter sequence consisting of p LSP parameters will be represented as θ[1], θ[2], . . . , θ[p]. “p” represents the order of prediction which is an integer equal to or greater than 1. The symbol in brackets ([ ]) represents index. For example, θ[i] indicates the ith LSP parameter in an LSP parameter sequence θ[1], θ[2], . . . , θ[p].
A symbol written in the upper right of θ in brackets indicates frame number. For example, an LSP parameter sequence generated for the sound signals in the fth frame is represented as θ^{[f]}[1], θ^{[f]}[2], . . . , θ^{[f]}[p]. However, since most processing is conducted within a frame in a closed manner, indication of the upper right frame number is omitted for parameters that correspond to the current frame (the fth frame). Omission of a frame number is intended to mean parameters generated for the current frame. That is, θ[i]=θ^{[f]}[i] holds.
A symbol written in the upper right without brackets represents exponentiation. That is, θ^{k}[i] means the kth power of θ[i].
Although symbols used in the text such as “˜”, “{circumflex over ( )}”, and “^{−}” should be originally indicated immediately above the following letter, they are indicated immediately before the corresponding letter due to limitations in text denotation. In mathematical expressions, such symbols are indicated at the appropriate position, namely immediately above the corresponding letter.
At step S100, a speech sound digital signal (hereinafter referred to as input sound signal) in the time domain per frame, which defines a predetermined time segment, is input to a conventional encoding apparatus 9. The encoding apparatus 9 performs processing in the processing units described below on the input sound signal on a perframe basis.
A perframe input sound signal is input to a linear prediction analysis unit 105, a feature amount extracting unit 120, a frequency domain encoding unit 150, and a time domain encoding unit 170.
At step S105, the linear prediction analysis unit 105 performs linear prediction analysis on the perframe input sound signal to determine a linear prediction coefficient sequence a[1], a[2], . . . , a[p], and outputs it. Here, a[i] is a linear prediction coefficient of the ith order. Each coefficient a[i] in the linear prediction coefficient sequence is coefficient a[i] (i=1, 2, . . . , p) that is obtained when input sound signal z is modeled with the linear prediction model represented by Formula (1):
The linear prediction coefficient sequence a[1], a[2], . . . , a[p] output by the linear prediction analysis unit 105 is input to an LSP generating unit 110.
At step S110, the LSP generating unit 110 determines and outputs a series of LSP parameters, θ[1], θ[2], . . . , θ[p], corresponding to the linear prediction coefficient sequence a[1], a[2], . . . , a[p] output from the linear prediction analysis unit 105. In the following description, the series of LSP parameters, θ[1], θ[2], . . . , θ[p], will be referred to as an LSP parameter sequence. The LSP parameter sequence θ[1], θ[2], . . . , θ[p] is a series of parameters that are defined as the root of the sum polynomial defined by Formula (2) and the difference polynomial defined by Formula (3).
F_{1}(z)=A(z)+z^{−(p+1)}A(z^{−1}) (2)
F_{2}(z)=A(z)−z^{−(p+1)}A(z^{−1}) (3)
The LSP parameter sequence θ[1], θ[2], . . . , θ[p] is a series in which values are arranged in ascending order. That is, it satisfies
0<θ[1]<θ[2]< . . . <θ[p]<π.
The LSP parameter sequence θ[1], θ[2], . . . , θ[p] output by the LSP generating unit 110 is input to an LSP encoding unit 115.
At step S115, the LSP encoding unit 115 encodes the LSP parameter sequence θ[1], θ[2], . . . , θ[p] output by the LSP generating unit 110, determines LSP code C1 and a quantized LSP parameter series {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] corresponding to the LSP code C1, and outputs them. In the following description, the quantized LSP parameter series {circumflex over ( )}[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] will be referred to as a quantized LSP parameter sequence.
The quantized LSP parameter sequence {circumflex over ( )}[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] output by the LSP encoding unit 115 is input to a quantized linear prediction coefficient generating unit 900, a delay input unit 165, and a time domain encoding unit 170. The LSP code C1 output by the LSP encoding unit 115 is input to an output unit 175.
At step S120, the feature amount extracting unit 120 extracts the magnitude of the temporal variation in the input sound signal as the feature amount. When the extracted feature amount is smaller than a predetermined threshold (i.e., when the temporal variation in the input sound signal is small), the feature amount extracting unit 120 implements control so that the quantized linear prediction coefficient generating unit 900 will perform the subsequent processing. At the same time, the feature amount extracting unit 120 inputs information indicating the frequency domain encoding method to the output unit 175 as identification code Cg. Meanwhile, when the extracted feature amount is equal to or greater than the predetermined threshold (i.e., when the temporal variation in the input sound signal is large), the feature amount extracting unit 120 implements control so that the time domain encoding unit 170 will perform the subsequent processing. At the same time, the feature amount extracting unit 120 inputs information indicating the time domain encoding method to the output unit 175 as identification code Cg.
Processes in the quantized linear prediction coefficient generating unit 900, a quantized linear prediction coefficient adjusting unit 905, an approximate smoothed power spectral envelope series calculating unit 910, and the frequency domain encoding unit 150 are executed when the feature amount extracted by the feature amount extracting unit 120 is smaller than the predetermined threshold (i.e., when the temporal variation in the input sound signal is small) (step S121).
At step S900, the quantized linear prediction coefficient generating unit 900 determines a series of linear prediction coefficients, {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p], from the quantized LSP parameter sequence {circumflex over ( )}[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] output by the LSP encoding unit 115, and outputs it. In the following description, the linear prediction coefficient series {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p] will be referred to as a quantized linear prediction coefficient sequence.
The quantized linear prediction coefficient sequence {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p] output by the quantized linear prediction coefficient generating unit 900 is input to the quantized linear prediction coefficient adjusting unit 905.
At step S905, the quantized linear prediction coefficient adjusting unit 905 determines and outputs a series {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }of value {circumflex over ( )}a[i]×(γR)^{i}, which is the product of the ithorder coefficient {circumflex over ( )}a[i] (i=1, . . . , p) in the quantized linear prediction coefficient sequence {circumflex over ( )}[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p] output by the quantized linear prediction coefficient generating unit 900 and the ith power of adjustment factor γR. Here, the adjustment factor γR is a predetermined positive integer equal to or smaller than 1. In the following description, the series {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }will be referred to as a adjusted quantized linear prediction coefficient sequence.
The adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }output by the quantized linear prediction coefficient adjusting unit 905 is input to the approximate smoothed power spectral envelope series calculating unit 910.
At step S910, using each coefficient {circumflex over ( )}a[i]×(γR)^{i }in the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }output by the quantized linear prediction coefficient adjusting unit 905, the approximate smoothed power spectral envelope series calculating unit 910 generates an approximate smoothed power spectral envelope series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] by Formula (4) and outputs it. Here, exp(⋅) is an exponential function whose base is Napier'"'"'s constant, j is the imaginary unit, and σ^{2 }is prediction residual energy.
As defined by Formula (4), the approximate smoothed power spectral envelope series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] is a frequencydomain series corresponding to the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p}.
The approximate smoothed power spectral envelope series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] output by the approximate smoothed power spectral envelope series calculating unit 910 is input to the frequency domain encoding unit 150.
In the following, the reason why a series of values defined by Formula (4) is called an approximate smoothed power spectral envelope series will be explained.
With a pthorder autoregressive process which is an allpole model, input sound signal x[t] at time t is represented by Formula (5) with its own values in the past back to time p, i.e., x[t−1], . . . , x[t−p], a prediction residual e[t], and linear prediction coefficients a[1], a[2], . . . , a[p]. Then, each coefficient W[n] (n=1, . . . , N) in a power spectral envelope series W[1], W[2], . . . , W[N] of the input sound signal is represented by Formula (6):
x[t]+a[1]x[t−1]+ . . . +a[p]x[t−p]=e[t] (5)
Here, a series W_{γR}[1], W_{γR}[2], . . . , W_{γR}[N] defined by
sin which a[i] in Formula (6) is replaced with a[i]×(γR)^{i }is equivalent to the power spectral envelope series W[1], W[2], . . . , W[N] of the input sound signal defined by Formula (6) but with the waves of the amplitude smoothed. In other words, processing for adjusting a linear prediction coefficient by multiplying linear prediction coefficient a[i] by the ith power of the adjustment factor γR is equivalent to processing that flats the waves of the amplitude of the power spectral envelope in the frequency domain (processing for smoothing the power spectral envelope). Accordingly, the series W_{γR}[1], W_{γR}[2], . . . , W_{γR}[N] defined by Formula (7) is called a smoothed power spectral envelope series.
The series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] defined by Formula (4) is equivalent to a series of approximations of the individual values in the smoothed power spectral envelope series W_{γR}[1], W_{γR}[2], . . . , W_{γR}[N] defined by Formula (7). Accordingly, the series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] defined by Formula (4) is called an approximate smoothed power spectral envelope series.
At step S150, the frequency domain encoding unit 150 normalizes each value X[n] (n=1, . . . , N) in a frequency domain signal sequence X[1], X[2], . . . , X[N], generated by converting the input sound signal into the frequency domain, with the square root of each value ˜W_{γR}[n] in the approximate smoothed power spectral envelope series, thereby determining a normalized frequency domain signal sequence X_{N}[1], X_{N}[2], . . . , X_{N}[N]. That is to say, X_{N}[n]=X[n]/sqrt (˜W_{γR}[n]) holds. Here, sqrt(y) represents the square root of y. The frequency domain encoding unit 150 then encodes the normalized frequency domain signal sequence X_{N}[1], X_{N}[2], . . . , X_{N}[N] by variable length encoding to generate frequency domain signal codes.
The frequency domain signal codes output by the frequency domain encoding unit 150 are input to the output unit 175.
The delay input unit 165 and the time domain encoding unit 170 are executed when the feature amount extracted by the feature amount extracting unit 120 is equal to or greater than the predetermined threshold (i.e., when the temporal variation in the input sound signal is large) (step S121).
At step S165, the delay input unit 165 holds the input quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], and outputs it to the time domain encoding unit 170 with a delay equivalent to the duration of one frame. For example, if the current frame is the fth frame, the quantized LSP parameter sequence for the f−1th frame, {circumflex over ( )}θ^{[f−1]}[1], {circumflex over ( )}θ^{[f−1]}[2], . . . , {circumflex over ( )}^{[f−1]}[p], is output to the time domain encoding unit 170.
At step S170, the time domain encoding unit 170 carries out encoding by determining a synthesized signal by applying the synthesis filter to a signal generated by synthesis of the waveforms contained in the adaptive codebook and the waveforms contained in the fixed codebook, and determining the indices for the respective codebooks so that the distortion between the synthesized signal determined and the input sound signal is minimized. When determining the indices for the codebooks so that the distortion between the synthesized signal and the input sound signal is minimized, the codebook indices are determined so as to minimize the value given by applying an auditory weighting filter to a signal representing the difference of the synthesized signal from the input sound signal. The auditory weighting filter is a filter for determining distortion when selecting the adaptive codebook and/or the fixed codebook.
The filter coefficients of the synthesis filter and the auditory weighting filter are generated by use of the quantized LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], and the quantized LSP parameter sequence for the f−1th frame, {circumflex over ( )}θ^{[f−1]}[1], {circumflex over ( )}^{[f−1]}[2], . . . , {circumflex over ( )}θ^{[f−1]}[p].
Specifically, a frame is first divided into two subframes, and the filter coefficients for the synthesis filter and the auditory weighting filter are determined as follows.
In the latterhalf subframe, each coefficient {circumflex over ( )}a[i] in a quantized linear prediction coefficient sequence {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p], which is a coefficient sequence obtained by converting the quantized LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], into linear prediction coefficients, is employed for the filter coefficient of the synthesis filter. For the filter coefficients of the auditory weighting filter, a series of values,
{circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p},
is employed which is determined by multiplying each coefficient {circumflex over ( )}a[i] in the quantized linear prediction coefficient sequence {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p] by the ith power of adjustment factor γR.
In the firsthalf subframe, each coefficient ˜a[i] in an interpolated quantized linear prediction coefficient sequence ˜a[1], ˜a[2], . . . , ˜a[p], which is a coefficient sequence obtained by converting an interpolated quantized LSP parameter sequence ˜θ[1], ˜θ[2], . . . , ˜θ[p] into linear prediction coefficients, is employed for the filter coefficient of the synthesis filter. The interpolated quantized LSP parameter sequence ˜θ[1], ˜θ[2], . . . , ˜θ[p] is a series of intermediate values between each value {circumflex over ( )}θ[i] in the quantized LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], and each value {circumflex over ( )}θ^{[f−1]}[i] in the quantized LSP parameter sequence for the f−1th frame, {circumflex over ( )}θ^{[f−1]}[1], {circumflex over ( )}θ^{[f−1]}[2], . . . , {circumflex over ( )}θ^{[f−1]}[p], namely a series of values obtained by interpolating between the values {circumflex over ( )}θ[i] and {circumflex over ( )}θ^{[f−1]}[i]. For the filter coefficients of the auditory weighting filter, a series of values,
˜a[1]×(γR), ˜a[2]×(γR)^{2}, . . . , ˜a[p]×(γR)^{p},
is employed which is determined by multiplying each coefficient ˜a[i] in the interpolated quantized linear prediction coefficient sequence ˜a[1], ˜a[2], . . . , ˜a[p] by the ith power of the adjustment factor γR.
This has the effect of smoothing the transition between a decoded sound signal and the decoded sound signal for the preceding frame generated in the decoding apparatus. Note that the adjustment factor γ used in the time domain encoding unit 170 is the same as the adjustment factor γ used in the approximate smoothed power spectral envelope series calculating unit 910.
At step S175, the encoding apparatus 9 transmits, by way of the output unit 175, the LSP code C1 output by the LSP encoding unit 115, the identification code Cg output by the feature amount extracting unit 120, and either the frequency domain signal codes output by the frequency domain encoding unit 150 or the time domain signal codes output by the time domain encoding unit 170, to the decoding apparatus.
Nonpatent Literature 1: 3rd Generation Partnership Project (3GPP), “Extended Adaptive MultiRateWideband (AMRWB+) codec; Transcoding functions”, Technical Specification (TS) 26.290, Version 10.0.0, 201103.
Nonpatent Literature 2: M. Neuendorf, et al., “MPEG Unified Speech and Audio CodingThe ISO/MPEG Standard for HighEfficiency Audio Coding of All Content Types”, Audio Engineering Society Convention 132, 2012.
The adjustment factor γR serves to achieve encoding with small distortion that takes the sense of hearing into account to an increased degree by flattening the waves of the amplitude of a power spectral envelope more for a higher frequency when eliminating the influence of the power spectral envelope from the input sound signal.
In order for the frequency domain encoding unit to achieve encoding with small distortion taking into account the sense of hearing, it is necessary for the approximate smoothed power spectral envelope series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N] to approximate the smoothed power spectral envelope W_{γR}[1], W_{γR}[2], . . . , W_{γR}[N] with high accuracy. Stated differently, assuming that
a_{γR}[i]=a[i]×(γR)^{i}(i=1, . . . , p),
it is desirable that the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }is a series that approximates the adjusted linear prediction coefficient sequence a_{γR}[1], a_{γR}[2], . . . , a_{γR}[p] with high accuracy.
However, the LSP encoding unit of a conventional encoding apparatus performs encoding processing so that the distortion between the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] and the LSP parameter sequence θ[1], θ[2], . . . , θ[p] is minimized. This means determining the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] so that a power spectral envelope that does not take the sense of hearing into account (i.e., that has not been smoothed with adjustment factor γR) is approximated with high accuracy. Consequently, the distortion between the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p }generated from the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] and the adjusted linear prediction coefficient sequence a_{γR}[1], a_{γR}[2], . . . , a_{γR}[p] is not minimized, leading to large encoding distortion in the frequency domain encoding unit.
An object of the present invention is to provide encoding techniques that selectively use frequency domain encoding and time domain encoding in accordance with the characteristics of the input sound signal and that are capable of reducing the encoding distortion in frequency domain encoding compared to conventional techniques, and also generating LSP parameters that correspond to quantized LSP parameters for the preceding frame and are to be used in time domain encoding, from linear prediction coefficients resulting from frequency domain encoding or coefficients equivalent to linear prediction coefficients, typified by LSP parameters. Another object of the present invention is to generate coefficients equivalent to linear prediction coefficients having varying degrees of smoothing effect from coefficients equivalent to linear prediction coefficients used, for example, in the abovedescribed encoding technique.
In order to attain the objects, a frequency domain parameter sequence generating method according to a first aspect of the invention, implemented by a frequency domain parameter sequence generating apparatus having processing circuitry.
The frequency domain parameter sequence generating method, includes, where p is an integer equal to or greater than 1, a linear prediction coefficient sequence which is obtained by linear prediction analysis of audio signals in a predetermined time segment as a[1], a[2], . . . , a[p], and ω[1], ω[2], . . . , ω[p] are a frequency domain parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], determining, by the processing circuitry, a converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p] using the frequency domain parameter sequence ω[1], ω[2], . . . , ω[p] as input in a parameter sequence conversion step. The processing circuitry determines a value of each converted frequency domain parameter ˜ω[i] (i=1, 2, . . . , p) in the converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p] through linear transformation which is based on a relationship of values between ω[i] and one or more frequency domain parameters adjacent to ω[i].
A frequency domain parameter sequence generating method according to a second aspect of the invention, implemented by a frequency domain parameter sequence generating apparatus having processing circuitry.
The frequency domain parameter sequence generating method includes, where p is an integer equal to or greater than 1, and a linear prediction coefficient sequence obtained by linear prediction analysis of audio signals in a predetermined time segment as a[1], a[2], . . . , a[p]; ω[1], ω[2], . . . , ω[p] is one of an LSP parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], an LSF parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], and a frequency domain parameter sequence which is derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p] and in which all of ω[1], ω[2], . . . , ω[p] are present from 0 to π and, when all of linear prediction coefficients contained in the linear prediction coefficient sequence are 0, ω[1], ω[2], . . . , ω[p] are present from 0 to π at equal intervals; and each γ1 and γ2 is a adjustment factor which is a positive constant equal to or smaller than 1, and K is a predetermined p×p band matrix in which diagonal elements and elements that neighbor the diagonal elements in row direction have nonzero values, generating, by the processing circuitry, a converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ω[p] defined by a following formula
A frequency domain parameter sequence generating method according to a third aspect of the invention, implemented by a frequency domain parameter sequence generating apparatus having processing circuitry.
The frequency domain parameter sequence generating method, includes, where p is an integer equal to or greater than 1, a linear prediction coefficient sequence which is obtained by linear prediction analysis of audio signals in a predetermined time segment as a[1], a[2], . . . , a[p], is one of an ISP parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], and an ISF parameter sequence derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p]; and each γ1 and γ2 is a adjustment factor which is a positive constant equal to or smaller than 1, and K is a predetermined p−1×p−1 band matrix in which diagonal elements and elements that neighbor the diagonal elements in row direction have nonzero values, generating, by the processing circuitry, a converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p−1] defined by a following formula
A decoding method according to a fourth aspect of the invention, implemented by a decoding apparatus having processing circuitry.
The decoding method, includes: decoding, by the processing circuitry, input adjusted LSP codes to obtain a decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γ}[1], {circumflex over ( )}θ_{γ}[2], . . . , {circumflex over ( )}θ_{γ}[p]; with the frequency domain parameter sequence ω[1], ω[2], . . . , ω[p] being the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γ}[1], {circumflex over ( )}θ_{γ}[2], . . . , {circumflex over ( )}θ_{γ}[p], executing, by the processing circuitry, the parameter sequence conversion step of the frequency domain parameter sequence generating method described in the first aspect to thereby generate the converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p] as a decoded approximate LSP parameter sequence {circumflex over ( )}θ_{app}[1], {circumflex over ( )}θ_{app}[2], . . . , {circumflex over ( )}θ_{app}[p]; calculating, by the processing circuitry, a decoded smoothed power spectral envelope series {circumflex over ( )}W_{γ}[1], {circumflex over ( )}W_{γ}[2], . . . , {circumflex over ( )}W_{γ}[N] based on the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γ}[1], {circumflex over ( )}θ_{γ}[2], . . . , {circumflex over ( )}θ_{γ}[p]; generating, by the processing circuitry, decoded sound signals using the frequency domain signal sequence resulting from decoding of input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over ( )}W_{γ}[1], {circumflex over ( )}W_{γ}[2], . . . , {circumflex over ( )}W_{γ}[N]; generating, by the processing circuitry, decoded sound signals using the frequency domain signal sequence resulting from decoding of the input frequency domain signal codes and the decoded smoothed power spectral envelope series {circumflex over ( )}W_{γ}[1], {circumflex over ( )}W_{γ}[2], . . . , {circumflex over ( )}W_{γ}[N]; decoding, by the processing circuitry, input LSP codes to obtain a decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p]; and decoding, by the processing circuitry, input time domain signal codes, and generating decoded sound signals by synthesizing the time domain signal codes using either the decoded LSP parameter sequence for the preceding time segment or the decoded approximate LSP parameter sequence for the preceding time segment, and the decoded LSP parameter sequence for the predetermined time segment.
According to the encoding techniques of the present invention, it is possible to reduce the encoding distortion in frequency domain encoding compared to conventional techniques, and also obtain LSP parameters that correspond to quantized LSP parameters for the preceding frame and are to be used in time domain encoding from linear prediction coefficients resulting from frequency domain encoding or coefficients equivalent to linear prediction coefficients, typified by LSP parameters. It is also possible to generate coefficients equivalent to linear prediction coefficients having varying degrees of smoothing effect from coefficients equivalent to linear prediction coefficients used in, for example, the abovedescribed encoding technique.
Embodiments of the present invention will be described below. In the drawings used in the description below, components having the same function or steps that perform the same processing are denoted with the same reference characters and repeated descriptions are omitted.
A encoding apparatus according to a first embodiment obtains, in a frame for which time domain encoding is performed, LSP codes by encoding LSP parameters that have been converted from linear prediction coefficients. In a frame for which frequency domain encoding is performed, the encoding apparatus obtains adjusted LSP codes by encoding adjusted LSP parameters that have been converted from adjusted linear prediction coefficients. When time domain encoding is to be performed in a frame following a frame for which frequency domain encoding was performed, linear prediction coefficients generated by inverse adjustment of linear prediction coefficients that correspond to LSP parameters corresponding to adjusted LSP codes are converted to LSPs, which are then used as LSP parameters in the time domain encoding for the following frame.
A decoding apparatus according to the first embodiment obtains, in a frame for which time domain decoding is performed, linear prediction coefficients that have been converted from LSP parameters resulting from decoding of LSP codes and uses them for time domain decoding. In a frame for which frequency domain decoding is performed, the decoding apparatus uses adjusted LSP parameters generated by decoding adjusted LSP codes for the frequency domain decoding. When time domain decoding is to be performed in a frame following a frame for which frequency domain decoding was performed, linear prediction coefficients generated by inverse adjustment of linear prediction coefficients that correspond to LSP parameters corresponding to the adjusted LSP codes are converted to LSPs, which are then used as LSP parameters in the time domain decoding for the following frame.
In the encoding and decoding apparatuses according the first embodiment, as illustrated in
<Encoding Apparatus>
As shown in
The encoding apparatus 1 is a specialized device build by incorporating special programs into a known or dedicated computer having a central processing unit (CPU), main memory (random access memory or RAM), and the like, for example. The encoding apparatus 1 performs various kinds of processing under the control of the central processing unit, for example. Data input to the encoding apparatus 1 or data resulting from various kinds of processing are stored in the main memory, for example, and data stored in the main memory are retrieved for use in other processing as necessary. At least some of the processing components of the encoding apparatus 1 may be implemented by hardware such as an integrated circuit.
As shown in
With the configuration of the first embodiment, when the feature amount extracted by the feature amount extracting unit 120 in the preceding frame was smaller than the predetermined threshold (i.e., when temporal variation in the input sound signal was small), the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] is not generated and thus cannot be input to the delay input unit 165. The quantized linear prediction coefficient inverse adjustment unit 155 and the inverseadjusted LSP generating unit 160 are processing components added for addressing this: when the feature amount extracted by the feature amount extracting unit 120 in the preceding frame was smaller than the predetermined threshold (i.e., when temporal variation in the input sound signal was small), they generate a series of approximations of the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] for the preceding frame to be used in the time domain encoding unit 170, from the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p]. In this case, an inverseadjusted LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] is the series of approximations of the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p].
<Encoding Method>
Referring to
At step S125, the linear prediction coefficient adjusting unit 125 determines a series of coefficient, a_{γR}[i]=a[i]×γR^{i}, which is the product of each coefficient a[i] (i=1, . . . , p) in the linear prediction coefficient sequence a[1], a[2], . . . , a[p] output by the linear prediction analysis unit 105 and the ith power of adjustment factor γR, and outputs it. In the following description, the series a_{γR}[1], a_{γR}[2], . . . , a_{γR}[p] determined will be called a adjusted linear prediction coefficient sequence.
The adjusted linear prediction coefficient sequence a_{γR}[1], a_{γR}[2], . . . , a_{γR}[p] output by the linear prediction coefficient adjusting unit 125 is input to the adjusted LSP generating unit 130.
At step S130, the adjusted LSP generating unit 130 determines and outputs a adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , θ_{γR}[p], which is a series of LSP parameters corresponding to the adjusted linear prediction coefficient sequence a_{γR}[1], a_{γR}[2], . . . , a_{γR}[p] output by the linear prediction coefficient adjusting unit 125. The adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , θ_{γR}[p] is a series in which values are arranged in ascending order. That is, it satisfies
0<θ_{γR}[1]<θ_{γR}[2]< . . . <θ_{γR}[p]<π.
The adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , θ_{γR}[p] output by the adjusted LSP generating unit 130 is input to the adjusted LSP encoding unit 135.
At step S135, the adjusted LSP encoding unit 135 encodes the adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , θ_{γR}[p] output by the adjusted LSP generating unit 130, and generates adjusted LSP code Cγ and a series of quantized adjusted LSP parameters, {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p], corresponding to the adjusted LSP code Cγ, and outputs them. In the following description, the series {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] will be called a adjusted quantized LSP parameter sequence.
The adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP encoding unit 135 is input to the quantized linear prediction coefficient generating unit 140. The adjusted LSP code Cγ output by the adjusted LSP encoding unit 135 is input to the output unit 175.
At step S140, the quantized linear prediction coefficient generating unit 140 generates and outputs a series of linear prediction coefficients, {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p], from the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP encoding unit 135. In the following description, the series {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}_{γR}[p] will be called a adjusted quantized linear prediction coefficient sequence.
The adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a_{γ}[1], {circumflex over ( )}a_{γ}[2], . . . , {circumflex over ( )}a_{γ}[p] output by the quantized linear prediction coefficient generating unit 140 is input to the first quantized smoothed power spectral envelope series calculating unit 145 and the quantized linear prediction coefficient inverse adjustment unit 155.
At step S145, the first quantized smoothed power spectral envelope series calculating unit 145 generates and outputs a quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] according to Formula (8) using each coefficient {circumflex over ( )}a_{γR}[i] in the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p] output by the quantized linear prediction coefficient generating unit 140.
The quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] output by the first quantized smoothed power spectral envelope series calculating unit 145 is input to the frequency domain encoding unit 150.
Processing in the frequency domain encoding unit 150 is the same as that performed by the frequency domain encoding unit 150 of the conventional encoding apparatus 9 except that it uses the quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] in place of the approximate smoothed power spectral envelope series ˜W_{γR}[1], ˜W_{γR}[2], . . . , ˜W_{γR}[N].
At step S155, the quantized linear prediction coefficient inverse adjustment unit 155 determines a series {circumflex over ( )}a_{γ}[1]/(γR), {circumflex over ( )}a_{γ}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γ}[p]/(γR)^{p }of value a_{γ}[i]/(γR)^{i }determined by dividing each value {circumflex over ( )}a_{γR}[i] in the adjusted quantized linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p] output by the quantized linear prediction coefficient generating unit 140 by the ith power of the adjustment factor γR, and outputs it. In the following description, the series {circumflex over ( )}a_{γ}[1]/(γR), {circumflex over ( )}a_{γ}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γ}[p]/(γR)^{p }will be called an inverseadjusted linear prediction coefficient sequence. The adjustment factor γR is set to the same value as the adjustment factor γR used in the linear prediction coefficient adjusting unit 125.
The inverseadjusted linear prediction coefficient sequence {circumflex over ( )}a_{γ}[1]/(γR), {circumflex over ( )}a_{γ}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γ}[p]/(γR)^{p }output by the quantized linear prediction coefficient inverse adjustment unit 155 is input to the inverseadjusted LSP generating unit 160.
At step S160, the inverseadjusted LSP generating unit 160 determines and outputs a series of LSP parameters, {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p], from the inverseadjusted linear prediction coefficient sequence {circumflex over ( )}a_{γ}[1]/(γR), {circumflex over ( )}a_{γ}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γ}[p]/(γR)^{p }output by the quantized linear prediction coefficient inverse adjustment unit 155. In the following description, the LSP parameter series {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ[p] will be called an inverseadjusted LSP parameter sequence. The inverseadjusted LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] is a series in which values are arranged in ascending order. That is, it is a series that satisfies
0<{circumflex over ( )}θ′[1]<{circumflex over ( )}θ′[2]< . . . <{circumflex over ( )}θ′[p]<π.
The inverseadjusted LSP parameters {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] output by the inverseadjusted LSP generating unit 160 are input to the delay input unit 165 as a quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p]. That is, the inverseadjusted LSP parameters {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] are used in place of the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p].
At step S175, the encoding apparatus 1 sends, by way of the output unit 175, the LSP code C1 output by the LSP encoding unit 115, the identification code Cg output by the feature amount extracting unit 120, the adjusted LSP code Cγ output by the adjusted LSP encoding unit 135, and either the frequency domain signal codes output by the frequency domain encoding unit 150 or the time domain signal codes output by the time domain encoding unit 170, to the decoding apparatus 2.
<Decoding Apparatus>
As illustrated in
The decoding apparatus 2 is a specialized device build by incorporating special programs into a known or dedicated computer having a central processing unit (CPU), main memory (random access memory or RAM), and the like, for example. The decoding apparatus 2 performs various kinds of processing under the control of the central processing unit, for example. Data input to the decoding apparatus 2 or data resulting from various kinds of processing are stored in the main memory, for example, and data stored in the main memory are retrieved for use in other processing as necessary. At least some of the processing components of the decoding apparatus 2 may be implemented by hardware such as an integrated circuit.
<Decoding Method>
Referring to
At step S200, a code sequence generated in the encoding apparatus 1 is input to the decoding apparatus 2. The code sequence contains the LSP code C1, identification code Cg, adjusted LSP code Cγ, and either frequency domain signal codes or time domain signal codes.
At step S205, the identification code decoding unit 205 implements control so that the adjusted LSP code decoding unit 215 will execute the subsequent processing if the identification code Cg contained in the input code sequence corresponds to information indicating the frequency domain encoding method, and so that the LSP code decoding unit 210 will execute the subsequent processing if the identification code Cg corresponds to information indicating the time domain encoding method.
The adjusted LSP code decoding unit 215, the decoded linear prediction coefficient generating unit 220, the first decoded smoothed power spectral envelope series calculating unit 225, the frequency domain decoding unit 230, the decoded linear prediction coefficient inverse adjustment unit 235, and the decoded inverseadjusted LSP generating unit 240 are executed when the identification code Cg contained in the input code sequence corresponds to information indicating the frequency domain encoding method (step S206).
At step S215, the adjusted LSP code decoding unit 215 obtains a decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] by decoding the adjusted LSP code Cγ contained in the input code sequence, and outputs it. That is, it obtains and outputs a decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] which is a sequence of LSP parameters corresponding to the adjusted LSP code Cγ. The same symbols are used because the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] obtained here is identical to the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] generated by the encoding apparatus 1 if the adjusted LSP code Cγ output by the encoding apparatus 1 is accurately input to the decoding apparatus 2 without being affected by code errors or the like.
The decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP code decoding unit 215 is input to the decoded linear prediction coefficient generating unit 220.
At step S220, the decoded linear prediction coefficient generating unit 220 generates and outputs a series of linear prediction coefficients, {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p], from the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP code decoding unit 215. In the following description, the series {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p] will be called a decoded adjusted linear prediction coefficient sequence.
The decoded linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p] output by the decoded linear prediction coefficient generating unit 220 is input to the first decoded smoothed power spectral envelope series calculating unit 225 and the decoded linear prediction coefficient inverse adjustment unit 235.
At step S225, the first decoded smoothed power spectral envelope series calculating unit 225 generates and outputs a decoded smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] according to Formula (8) using each coefficient {circumflex over ( )}a_{γR}[i] in the decoded adjusted linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[2], . . . , {circumflex over ( )}a_{γR}[p] output by the decoded linear prediction coefficient generating unit 220.
The decoded smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] output by the first decoded smoothed power spectral envelope series calculating unit 225 is input to the frequency domain decoding unit 230.
At step S230, the frequency domain decoding unit 230 decodes the frequency domain signal codes contained in the input code sequence to determine a decoded normalized frequency domain signal sequence X_{N}[1], X_{N}[2], . . . , X_{N}[N]. Next, the frequency domain decoding unit 230 obtains a decoded frequency domain signal sequence X[1], X[2], . . . , X[N] by multiplying each value X_{N}[n] (n=1, . . . , N) in the decoded normalized frequency domain signal sequence X_{N}[1], X_{N}[2], . . . , X_{N}[N] by the square root of each value {circumflex over ( )}W_{γR}[n] in the decoded smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N], and outputs it. That is, it calculates X[n]=X_{N}[n]×sqrt({circumflex over ( )}W_{γR}[n]). It then converts the decoded frequency domain signal sequence X[1], X[2], . . . , X[N] into the time domain to obtain and output decoded sound signals.
At step S235, the decoded linear prediction coefficient inverse adjustment unit 235 determines and outputs a series, {circumflex over ( )}a_{γR}[1]/(γR), {circumflex over ( )}a_{γR}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γR}[p]/(γR)^{p}, of value {circumflex over ( )}a_{γ}[i]/(γR)^{i }by dividing each value {circumflex over ( )}a_{γR}[i] in the decoded adjusted linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1], {circumflex over ( )}a_{γR}[^{2}], . . . , {circumflex over ( )}a_{γR}[p] output by the decoded linear prediction coefficient generating unit 220 by the ith power of the adjustment factor γR. In the following description, the series {circumflex over ( )}a_{γR}[1]/(γR), {circumflex over ( )}a_{γR}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γR}[p]/(γR)^{p }will be called a decoded inverseadjusted linear prediction coefficient sequence. The adjustment factor γR is set to the same value as the adjustment factor γR used in the linear prediction coefficient adjusting unit 125 of the encoding apparatus 1.
The decoded inverseadjusted linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1]/(γR), {circumflex over ( )}a_{γR}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γR}[p]/(γR)^{p }output by the decoded linear prediction coefficient inverse adjustment unit 235 is input to the decoded inverseadjusted LSP generating unit 240.
At step S240, the decoded inverseadjusted LSP generating unit 240 determines an LSP parameter series {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] from the decoded inverseadjusted linear prediction coefficient sequence {circumflex over ( )}a_{γR}[1]/(γR), {circumflex over ( )}a_{γR}[2]/(γR)^{2}, . . . , {circumflex over ( )}a_{γR}[p]/(γR)^{p}, and outputs it. In the following description, the LSP parameter series {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] will be called a decoded inverseadjusted LSP parameter sequence.
The decoded inverseadjusted LSP parameters {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] output by the decoded inverseadjusted LSP generating unit 240 are input to the delay input unit 245 as a decoded LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p].
The LSP code decoding unit 210, the delay input unit 245, and the time domain decoding unit 250 are executed when the identification code Cg contained in the input code sequence corresponds to information indicating the time domain encoding method (step S206).
At step S210, the LSP code decoding unit 210 decodes the LSP code C1 contained in the input code sequence to obtain a decoded LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p], and outputs it. That is, it obtains and outputs a decoded LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p], which is a sequence of LSP parameters corresponding to the LSP code C1.
The decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] output by the LSP code decoding unit 210 is input to the delay input unit 245 and the time domain decoding unit 250.
At step S245, the delay input unit 245 holds the input decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] and outputs it to the time domain decoding unit 250 with a delay equivalent to the duration of one frame. For instance, if the current frame is the fth frame, the decoded LSP parameter sequence for the f−1th frame, {circumflex over ( )}θ^{f−1}[1], {circumflex over ( )}θ^{f−1}[2], . . . , {circumflex over ( )}θ^{f−1}[p], is output to the time domain decoding unit 250.
When the identification code Cg contained in the input code corresponds to information indicating the frequency domain encoding method, the decoded inverseadjusted LSP parameter sequence {circumflex over ( )}θ′[1], {circumflex over ( )}θ′[2], . . . , {circumflex over ( )}θ′[p] output by the decoded inverseadjusted LSP generating unit 240 is input to the delay input unit 245 as the decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p].
At step S250, the time domain decoding unit 250 identifies the waveforms contained in the adaptive codebook and waveforms in the fixed codebook from the time domain signal codes contained in the input code sequence. By applying the synthesis filter to a signal generated by synthesis of the waveforms in the adaptive codebook and the waveforms in the fixed codebook that have been identified, a synthesized signal from which the effect of the spectral envelope has been removed is determined, and the synthesized signal determined is output as a decoded sound signal.
The filter coefficients for the synthesis filter are generated using the decoded LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], and the decoded LSP parameter sequence for the f−1th frame, {circumflex over ( )}θ^{f−1}[1], {circumflex over ( )}θ^{f−1}[2], . . . , {circumflex over ( )}θ^{f−1}[p].
Specifically, a frame is first divided into two subframes, and the filter coefficients for the synthesis filter are determined as follows.
In the latterhalf subframe, a series of values
{circumflex over ( )}a[1]×(γR), {circumflex over ( )}a[2]×(γR)^{2}, . . . , {circumflex over ( )}a[p]×(γR)^{p}
is used as filter coefficients for the synthesis filter. This is obtained by multiplying each coefficient {circumflex over ( )}a[i] of the decoded linear prediction coefficients {circumflex over ( )}a[1], {circumflex over ( )}a[2], . . . , {circumflex over ( )}a[p], which is a coefficient sequence generated by converting the decoded LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], into linear prediction coefficients, by the ith power of the adjustment factor γR.
In the firsthalf subframe, a series of values
˜a[1]×(γR), ˜a[2]×(γR)^{2}, . . . , ˜a[p]×(γR)^{p}
which is obtained by multiplying each coefficient ˜a[i] of decoded interpolated linear prediction coefficients ˜a[1], ˜a[2], . . . , ˜a[p] by the ith power of the adjustment factor γR, is used as filter coefficients for the synthesis filter. The decoded interpolated linear prediction coefficients ˜a[1], ˜a[2], . . . , ˜a[p] is a coefficient sequence generated by converting, into linear prediction coefficients, the decoded interpolated LSP parameter sequence ˜θ[1], ˜θ[2], . . . , ˜θ[p], which is a series of intermediate values between each value {circumflex over ( )}θ[i] in the decoded LSP parameter sequence for the fth frame, {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], and each value {circumflex over ( )}θ^{[f−1]}[i] in the decoded LSP parameter sequence for the f−1th frame, θ^{[f−1]}[1], θ^{[f−1]}[2], . . . , θ^{[f−1]}[p]. That is,
˜θ[i]=0.5 ×{circumflex over ( )}θ^{[f−1]}[i]+0.5×^θ[i] (i=1, . . . , p).
<Effects of the First Embodiment>
The adjusted LSP encoding unit 135 of the encoding apparatus 1 determines such a adjusted quantized LSP parameter sequence ^θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] that minimizes the quantizing distortion between the adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , θ_{γR}[p] and the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p]. This can determine the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] so that a power spectral envelope series that takes into account the sense of hearing (i.e., that has been smoothed with adjustment factor γR) is approximated with high accuracy. The quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N], which is a power spectral envelope series obtained by expanding the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] into the frequency domain, can approximate the smoothed power spectral envelope series W_{γR}[1], W_{γR}[2], . . . , W_{γR}[N] with high accuracy. When the code amount of the LSP code C1 is the same as that of the adjusted LSP code Cγ, the first embodiment yields smaller encoding distortion in frequency domain encoding than the conventional technique. In addition, assuming an equal encoding distortion to that in the conventional encoding method, the adjusted LSP code Cγ achieves a further smaller code amount compared to the conventional method than the LSP code C1 does. Thus, with a encoding distortion equal to that in the conventional method, the code amount can be reduced compared to the conventional method, whereas with the same code amount as the conventional method, encoding distortion can be reduced compared to the conventional method.
The encoding apparatus 1 and decoding apparatus 2 of the first embodiment are expensive in terms of calculation in the inverseadjusted LSP generating unit 160 and the decoded inverseadjusted LSP generating unit 240 in particular. To address this, a encoding apparatus 3 in a second embodiment directly generates an approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app}, which is a series of approximations of the values in the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], from the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] without the intermediation of linear prediction coefficients. Similarly, a decoding apparatus 4 in the second embodiment directly generates a decoded approximate LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app}, which is a series of approximations of the values in the decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], from the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] without the intermediation of linear prediction coefficients.
<Encoding Apparatus>
The encoding apparatus 3 differs from the encoding apparatus 1 of the first embodiment in that it does not include the quantized linear prediction coefficient inverse adjustment unit 155 and the inverseadjusted LSP generating unit 160 but includes an LSP linear transformation unit 300 instead.
Utilizing the nature of LSP parameters, the LSP linear transformation unit 300 applies approximate linear transformation to a adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] to generate an approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app}.
First, the nature of LSP parameters will be described.
Although the LSP linear transformation unit 300 applies approximate transformation to a series of quantized LSP parameters, the nature of an unquantized LSP parameter sequence will be discussed first because the nature of a quantized LSP parameter series is basically the same as the nature of an unquantized LSP parameter sequence.
An LSP parameter sequence θ[1], θ[2], . . . , θ[p] is a parameter sequence in the frequency domain that is correlated with the power spectral envelope of the input sound signal. Each value in the LSP parameter sequence is correlated with the frequency position of the extreme of the power spectral envelope of the input sound signal. The extreme of the power spectral envelope is present at a frequency position between θ[i] and θ[i+1]; and with a steeper slope of a tangent around the extreme, the interval between θ[i] and θ[i+1] (i.e., the value of θ[i+1]−θ[i]) becomes smaller. In other words, as the height difference in the waves of the amplitude of the power spectral envelope is larger, the interval between θ[i] and θ[i+1] becomes less even for each i (i=1, 2, . . . , p−1). Conversely, when there is almost no height difference in the waves of the power spectral envelope, the interval between θ[i] and θ[i+1] is close to an equal interval for each value of i.
As the value of the adjustment factor γ becomes smaller, the height difference in the waves of the amplitude of smoothed power spectral envelope series W_{γ}[1], W_{γ}[2], . . . , W_{γ}[N], defined by Formula (7), becomes smaller than the height difference in the waves of the amplitude of the power spectral envelope series W[1], W[2], . . . , W[N] defined by Formula (6). It can be accordingly said that a smaller value of the adjustment factor γ makes the interval between θ[i] and θ[i+1] closer to an equal interval. When γ has no influence (i.e., γ=0), this corresponds to the case of a flat power spectral envelope.
When the adjustment factor γ=0, adjusted LSP parameters θ_{γ=0}[1], θ_{γ=0}[2], . . . , θ_{γ=0}[p] are
in which case the interval between θ[i] and θ[i+1] is equal for all i=1, . . . , p31 1. When γ=1, the adjusted LSP parameter sequence θ_{γ=1}[1], θ_{γ=1}[2], . . . , θ_{γ=1}[p] and the LSP parameter sequence θ[1], θ[2], . . . , θ[p] are equivalent. The adjusted LSP parameters satisfy the property:
0<θ_{65}[1]<θ_{γ}[2] . . . <θ_{65 }[p]<π.
As shown in
when
θ_{γ1}[i]−θ_{γ1}[i−1]>θ_{γ1}[i+1]−θ_{γ1}[i] (9)
then the following properties hold:
θ_{γ2}[i+1]−θ_{γ2}[i]<θ_{γ1}[i+1]−θ_{γ1}[i], and
θ_{γ2}[i]−θ_{γ2}[i−1]>θ_{γ1}[i]−θ_{γ1}[i−1] (10)
When
θ_{γ1}[i]−θ_{γ1}[i−1]<θ_{γ1}[i+1]−θ_{γ1}[i] (11)
then the following properties hold:
θ_{γ2}[i+1]−θ_{γ2}[i]>θ_{γ1}[i+1]−θ_{γ1}[i], and
θ_{γ2}[i]−θ_{γ2}[i−1]<θ_{γ1}[i]−θ_{γ1}[i−1]. (12)
Formulas (9) and (10) indicate that when θ_{γ1}[i] is closer to θ_{γ1}[i+1] with respect to the midpoint between θ_{γ1}[i+1] and θ_{γ1}[i−1], θ_{γ}[i] will assume a value that is further closer to θ_{γ2}[i+1] (see
Formulas (11) and (12) indicate that when θ_{γ1}[i] is closer to θ_{γ1}[i−1] with respect to the midpoint between θ_{γ1}[i+1] and θ_{γ1}[i−1], θ_{γ}[i] will assume a value that is further closer to θ_{γ2}[i−1]. This means that on a twodimensional plane with the horizontal axis being the γ value and the vertical axis being the LSP parameter value, the slope of straight line connecting the point (γ1, θ_{γ1}[i]) and the point (γ2, θ_{γ2}[i]) is smaller than the slope of a straight line connecting the point (0, θ_{γ=0}[i]) and the point (γ1, θ_{γ1}[i]).
Based on the properties above, the relationship between θ_{γ1}[1], θ_{γ1}[2], . . . , θ_{γ1}[p] and θ_{γ2}[1], θ_{γ2}[2], . . . , θ_{γ2}[p] can be modeled with Formula (13), where Θ_{γ1}=(θ_{γ1}[1], θ_{γ1}[2], . . . , θ_{γ1}[p])^{T }and Θ_{γ2}=(θ_{γ2}[1], θ_{γ2}[2], . . . , θ_{γ2}[p])^{T}:
Θ_{γ2}≈K(Θ_{γ1}−Θ_{γ=0})(γ_{2}−γ_{1})+Θ_{γ1} (13)
where K is a p×p matrix defined by Formula (14).
In this case, 0<γ1, γ2≤1, and γ1≠γ2 hold. Although Formulas (9) to (12) describe the relationships on the assumption of γ1<γ2, the model of Formula (13) has no limitation on the relation of magnitude between γ1 and γ2; they may be either γ1<γ2 or γ1>γ2.
The matrix K is a band matrix that has nonzero values only in the diagonal components and elements adjacent to them and is a matrix representing the correlations described above that hold between LSP parameters corresponding to the diagonal components and the neighboring LSP parameters. Note that although Formula (14) illustrates a band matrix with a band width of three, the band width is not limited to three.
Assuming that
{tilde over (Θ)}_{γ2}=K(Θ_{γ1}−Θ_{γ=0})(γ_{2}−γ_{1})+Θ_{γ1} (13a),
then
˜Θ_{γ2}=(˜θ_{γ2}[1], ˜θ_{γ2}[2], . . . , ˜θ_{γ2}[p])^{T}
is an approximation of Θ_{γ2}.
Expanding Formula (13a) gives Formula (15) below:
{tilde over (θ)}_{γ2}[i]=z_{i}(θ_{γ1}[i−1]−θ_{γ=0}[i−1])+γ_{i}(θ_{γ1}[i+1]−θ_{γ=0}[i+1])+x_{i}(θ_{γ1}[i]−θ_{γ=0}[i])+θ_{γ1}[i] (15)
where i=2, . . . , p−1.
On a twodimensional plane with the horizontal axis representing the γ value and the vertical axis representing the LSP parameter value, let ^{−}θ_{γ2}[i] denote the value on the vertical axis corresponding to γ2 on an extension of straight line L1 that connects between the point (γ1, θ_{γ1}[i]) and the point (0, θ_{γ=0}[i]), namely the value on the vertical axis corresponding to γ2 as approximated by straight line approximation from the slope of straight line L1 connecting θ_{γ1}[i] and θ_{γ=0}[i] (see
holds. When γ1>γ2, it means straight line interpolation, while when γ1<γ2, it means straight line extrapolation.
In Formula (14), given that
then ˜θ_{γ2}[i]=^{−θ}_{γ2}[i], and ˜θ_{γ2}[i] obtained with the model of Formula (13a) matches the estimation ^{−}θ_{γ2}[i] of the LSP parameter value corresponding to γ2 as approximated by straight line approximation with a straight line that connects the point (γ1, θ_{γ1}[i]) and the point (0, θ_{γ=0}[i]) on the twodimensional plane. 101351 Given that u_{i }and v_{i }are positive values equal to or smaller than 1, assuming
in the Formula (14) above, Formula (15) can be rewritten as:
Formula (17) means adjusting the value of ^{−}θ_{γ2}[i] by weighting the differences between the ith LSP parameter θ_{γ1}[i] in the LSP parameter sequence, θ_{γ1}[1], θ_{γ1}[2], . . . , θ_{γ1}[p], and its preceding and following LSP parameter values (i.e., θ_{γ1}[i]−θ_{γ1}[i−1] and θ_{γ1}[i+1]−θ_{γ1}[i]) to obtain ˜θ_{γ2}[i]. That is to say, correlations such as shown in Formulas (9) through (12) above are reflected in the elements in the band portion (nonzero elements) of the matrix K in Formula (13a).
The values ˜θ_{γ2}[1], ˜θ_{γ2}[2], . . . , ˜θ_{γ2}[p] given by Formula (13a) are approximate values (estimated values) of LSP parameter values θ_{γ2}[1], θ_{γ2}[2], . . . , θ_{γ2}[p] when the linear prediction coefficient sequence a[1]×(γ2), . . . , a[p]×(γ2)^{p }is converted to LSP parameters.
Especially when γ2>γ1, the matrix K in Formula (14) tends to have positive values in the diagonal components and negative values in elements in the vicinity of them, as indicated by Formulas (16) and (17).
The matrix K is a preset matrix, which is prelearned using learning data, for example. How to learn the matrix K will be discussed later.
Similar properties also apply to quantized LSP parameters. That is, vectors Θ_{γ1 }and Θ_{γ2 }in the LSP parameter sequence in Formula (13) can be replaced with the vectors {circumflex over ( )}Θ_{γ1 }and {circumflex over ( )}Θ_{γ2 }in the quantized LSP parameter sequence, respectively. Specifically, {circumflex over ( )}Θ_{γ1}=({circumflex over ( )}θ_{γ1}[1], {circumflex over ( )}θ_{γ1}[2], . . . , {circumflex over ( )}θ_{γ1}[p])^{T }and {circumflex over ( )}Θ_{γ2}=({circumflex over ( )}θ_{γ2}[1], {circumflex over ( )}θ_{γ2}[2], . . . , {circumflex over ( )}θ_{γ2}[p])^{T}, then the following formula holds:
{circumflex over (Θ)}_{γ2}≈K({circumflex over (Θ)}_{γ1}−{circumflex over (Θ)}_{γ=0})(γ_{2}−γ_{1})+{circumflex over (Θ)}_{γ1} (13b).
Since matrix K is a band matrix, calculation cost required for calculating Formulas (13), (13a), and (13b) is very small.
The LSP linear transformation unit 300 included in the encoding apparatus 3 of the second embodiment generates an approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }from the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] based on Formula (13b). Note that the adjustment factor γR used in generation of the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] is the same as the adjustment factor γR used in the linear prediction coefficient adjusting unit 125.
<Encoding Method>
Referring to
Processing performed in the adjusted LSP encoding unit 135 is the same as the first embodiment. However, the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP encoding unit 135 is also input to the LSP linear transformation unit 300 in addition to the quantized linear prediction coefficient generating unit 140.
The LSP linear transformation unit 300, given {circumflex over ( )}Θ_{γ1}=({circumflex over ( )}θ_{γ1}[1], {circumflex over ( )}θ_{γ1}[2], . . . , {circumflex over ( )}θ_{γ1}[p])^{T}, determines and outputs an approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }according to
That is, using Formula (13b), the LSP linear transformation unit 300 determines a series of approximations, {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app}, of the quantized LSP parameter sequence. As γ1 and γ2 are constants, matrix K′ which is generated by multiplying the individual elements of matrix K by (γ2−γ1) may be used instead of the matrix K of Formula (18), and the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }may also be determined by
The approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }output by the LSP linear transformation unit 300 is input to the delay input unit 165 as the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p]. That is to say, in the time domain encoding unit 170, when the feature amount extracted by the feature amount extracting unit 120 for the preceding frame is smaller than the predetermined threshold (i.e., when temporal variation in the input sound signal was small, that is, when encoding in the frequency domain was performed), the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }for the preceding frame is used in place of the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] for the preceding frame.
<Decoding Apparatus>
The decoding apparatus 4 differs from the decoding apparatus 2 in the first embodiment in that it does not include the decoded linear prediction coefficient inverse adjustment unit 235 and the decoded inverseadjusted LSP generating unit 240 but includes a decoded LSP linear transformation unit 400 instead.
<Decoding Method>
Referring to
Processing in the adjusted LSP code decoding unit 215 is the same as the first embodiment. However, the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP code decoding unit 215 is also input to the decoded LSP linear transformation unit 400 in addition to the decoded linear prediction coefficient generating unit 220.
The decoded LSP linear transformation unit 400 determines a decoded approximate LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }according to Formula (18) with {circumflex over ( )}Θ_{γ1}=({circumflex over ( )}_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p])^{T}, and outputs it. That is, Formula (13b) is used to determine a series of approximations, {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app}, of the decoded LSP parameter sequence. As with the LSP linear transformation unit 300, the decoded approximate LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }may be determined by use of Formula (18a).
The decoded approximate LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }output by the decoded LSP linear transformation unit 400 is input to the delay input unit 245 as a decoded LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p]. It means that in the time domain decoding unit 250, when the identification code Cg for the preceding frame corresponds to information indicating the frequency domain encoding method, the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }for the preceding frame is used in place of the decoded LSP parameter sequence {circumflex over ( )}[1], {circumflex over ( )}[2], . . . , {circumflex over ( )}θ[p] for the preceding frame.
<Learning Process for Transformation Matrix K>
The transformation matrix K used in the LSP linear transformation unit 300 and the decoded LSP linear transformation unit 400 is determined in advance through the following process and prestored in storages (not shown) of the encoding apparatus 3 and the decoding apparatus 4.
(Step 1) For prepared sample data for speech sound signals corresponding to M frames, each sample data is subjected to linear prediction analysis to obtain linear prediction coefficients. A linear prediction coefficient sequence produced by linear prediction analysis of the mth (1≤m≤M) sample data is represented as a^{(m)}[1], a^{(m)}[2], . . . , a^{(m)}[p], and referred to as a linear prediction coefficient sequence a^{(m)}[1], a^{(m)}[2], . . . , a^{(m)}[p] corresponding to the mth sample data.
(Step 2) For each m, LSP parameters θ_{γ=1}^{(m)}[1], θ_{γ=1}^{(m)}[2], . . . , θ_{γ=1}^{(m)}[p] are determined from the linear prediction coefficient sequence a^{(m)}[1], a^{(m)}[2], . . . , a^{(m)}[p]. The LSP parameters θ_{γ=1}^{(m)}[1], θ_{γ=1}^{(m)}[2], . . . , θ_{γ=1}^{(m)}[p] are coded in a similar manner to the LSP encoding unit 115, thereby generating a quantized LSP parameter sequence {circumflex over ( )}θ_{γ=1}^{(m)}[1], {circumflex over ( )}θ_{γ=1}^{(m)}[2], . . . , {circumflex over ( )}θ_{γ=1}^{(m)}[p] . Here,
{circumflex over ( )}Θ^{(m)}_{γ1}=({circumflex over ( )}θ_{γ=1}^{(m)}[1], . . . , {circumflex over ( )}θ_{γ=1}^{(m)}[p])^{T}.
(Step 3) For each m, setting γL as a predetermined positive constant smaller than 1 (for example, γL=0.92), a adjusted linear prediction coefficient,
a_{γ}^{(m)}[i]=a^{(m)}[i]×(γL)^{i}
is calculated.
(Step 4) For each m, a adjusted LSP parameter sequence θ_{γL}^{(m)}[1], . . . , θ_{γL}^{(m)}[p] is determined from the adjusted linear prediction coefficient sequence a_{γL}^{(m)}[1], . . . , a_{γL}^{(m)}[p]. The adjusted LSP parameter sequence θ_{γL}^{(m)}[1], . . . , θ_{γL}^{(m)}[p] is coded in a similar manner to the adjusted LSP encoding unit 135, thereby generating a quantized LSP parameter sequence {circumflex over ( )}θ_{γL}^{(m)}[1], {circumflex over ( )}θ_{γL}^{(m)}[p]. Here,
{circumflex over ( )}Θ^{(m)}_{γ2}=({circumflex over ( )}θ_{γL}^{(m)}[1], . . . , {circumflex over ( )}θ_{γL}^{(m)}[p])^{T}.
Through Steps 1 to 4, M pairs of quantized LSP parameter sequences ({circumflex over ( )}Θ^{(m)}_{≢1}, {circumflex over ( )}Θ^{(m)}_{γ2}) are obtained. This set is used as learning data set Q, where Q={({circumflex over ( )}Θ^{(m)}_{γ1}, {circumflex over ( )}Θ^{(m)}_{γ2})m=1, . . . , M}. Note that all of the values of adjustment factor γL used in generation of the learning data set Q are common fixed values.
(Step 5) Each pair of LSP parameter sequences ({circumflex over ( )}Θ^{(m)}_{γ1}, {circumflex over ( )}Θ^{(m)}_{γ2}) contained in the learning data Q is substituted into the model of Formula (13b), where γ1=γL, γ2=1, {circumflex over ( )}Θ_{γ1}={circumflex over ( )}Θ^{(m)}_{γ1}, and {circumflex over ( )}Θ_{γ2}={circumflex over ( )}Θ^{(m)}_{γ2}, and the coefficients for matrix K are learned with the square error criterion. That is, a vector in which the components in the band portion of the matrix K are arranged in order from the top is defined as:
and B is obtained by
Learning of the matrix K is performed with the value of γL fixed. However, the matrix K used in the LSP linear transformation unit 300 does not have to be one that has been learned using the same value as the adjustment factor γR used in the encoding apparatus 3.
By way of example, values obtained by multiplying (γ2−γ1) and the elements in the band portion of the matrix K generated by the abovedescribed method given that p=15 and γL=0.92, namely the values of the elements in the band portion of matrix K′, are shown below. That is, the products of the values x_{1}, x_{2}, . . . , x_{15}, y_{1}, y_{2}, . . . , y_{14}, z_{2}, z_{3}, . . . , z_{15 }in Formula (14) and γ2−γ1 are xx_{1}, xx_{2}, . . . , xx_{15}, yy_{1}, yy_{2}, . . . , yy_{14}, zz_{2}, zz_{3}, . . . , zz_{15 }below:
 xx1=1.11499, yy1=−0.54272,
 zz2=−0.83414f, xx2=1.59810f, yy2=−0.70966,
 zz3=−0.49432, xx3=1.38370, yy3=−0.78076,
 zz4=−0.39319, xx4=1.23032, yy4=−0.67921,
 zz5=−0.39166, xx5=1.18521, yy5=−0.69088,
 zz6=−0.34784, xx6=1.04839, yy6=−0.60619,
 zz7=−0.41279, xx7=1.13305, yy7 =−0.63247,
 zz8=−0.36450, xx8=0.95694, yy8=−0.53039,
 zz9=−0.43984, xx9=1.01910, yy9=−0.51707,
 zz10=−0.40120, xx10=0.90395, yy10=−0.44594,
 zz11=−0.49262, xx11=1.07345, yy11=−0.51892,
 zz12=−0.41695, xx12=0.96596, yy12=−0.49247,
 zz13=−0.45002, xx13=1.00336, yy13=−0.48790,
 zz14=−0.46854, xx14=0.93258, yy14=−0.41927,
 zz15=−0.45020, xx15=0.88783.
When γ2>γ1 as in the above example, in which γ1=γL=0.92 and γ2=1, the diagonal components of matrix K′ assume values close to 1 as in the above example, while components neighboring the diagonal component assume negative values.
Conversely, when γ1>γ2, the diagonal components of matrix K′ assume negative values as in the example shown below, while components neighboring the diagonal component assume positive values. Values obtained by multiplying (γ2−γ1) and the elements in the band portion of the matrix K with p=15, γ1=1, and γ2=γL=0.92, namely the values of the elements in the band portion of matrix K′ can be as below, for example:
 xx1=−0.557012055, yy1=0.213853042,
 zz2=0.110112745, xx2 =−0.534830085, yy2=0.2440903,
 zz3=0.149879603, xx3=−0.522734808, yy3=0.23494022,
 zz4=0.144479327, xx4=−0.533013231, yy4=0.259021145,
 zz5=0.136523255, xx5=−0.502606738, yy5=0.248139539,
 zz6=0.138005088, xx6=−0.478327709, yy6=0.244219107,
 zz7=0.133771751, xx7=−0.467186849, yy7=0.243988642,
 zz8=0.13667916, xx8=−0.408737408, yy8=0.192803054,
 zz9=0.160602461, xx9=−0.427436157, yy9=0.190554547,
 zz10=0.147621742, xx10=−0.383087812, yy10=0.165954888,
 zz11=0.18358465, xx11=−0.434034351, yy11=0.183004742,
 zz12=0.166249458, xx12=−0.409482196, yy12=0.170107295,
 zz13=0.162343147, xx13=−0.409804718, yy13=0.165221097,
 zz14=0.178158258, xx14−−0.400869431, yy14=0.123020055,
 zz15=0.171958144, xx15=−0.447472325.
When γ1>y2, this corresponds to a case where {circumflex over ( )}Θ^{(m)}_{γ1 }is set as
{circumflex over ( )}Θ^{(m)}_{γ1}=({circumflex over ( )}θ_{γL}^{(m)}[1], . . . , {circumflex over ( )}θ_{γL}^{(m)}[p])^{T}
in Step 2 of <Learning Process for Transformation Matrix K>, {circumflex over ( )}Θ^{(m)}_{γ2 }is set as
{circumflex over ( )}Θ^{(m)}_{γ2}=({circumflex over ( )}θ_{γ=1}^{(m)}[1], . . . , {circumflex over ( )}θ_{γ=1}^{(m)}[p])^{T}
in Step 4, and each pair of LSP parameter sequences ({circumflex over ( )}Θ^{(m)}_{γ1}, {circumflex over ( )}Θ^{(m)}_{γ2}) contained in learning data Q is substituted into the model of Formula (13b) with γ1=1, γ2=γL, {circumflex over ( )}Θ_{γ1}={circumflex over ( )}Θ^{(m)}_{γ1}, and {circumflex over ( )}Θ_{γ2}={circumflex over ( )}Θ^{(m)}_{γ2 }in Step 5 and the coefficients for matrix K are learned with the square error criterion.
<Effects of the Second Embodiment>
The encoding apparatus 3 according to the second embodiment provides similar effects to the encoding apparatus 1 in the first embodiment because, as with the first embodiment, it has a configuration in which the quantized linear prediction coefficient generating unit 900, the quantized linear prediction coefficient adjusting unit 905, and the approximate smoothed power spectral envelope series calculating unit 910 of the conventional encoding apparatus 9 are replaced with the linear prediction coefficient adjusting unit 125, adjusted LSP generating unit 130, adjusted LSP encoding unit 135, quantized linear prediction coefficient generating unit 140, and the first quantized smoothed power spectral envelope series calculating unit 145. That is, when the encoding distortion is equal to that in a conventional method, the code amount can be reduced compared to the conventional method, whereas when the code amount is the same as in the conventional method, encoding distortion can be reduced compared to the conventional method.
In addition, the calculation cost of the encoding apparatus 3 in the second embodiment is low because K is a band matrix in calculation of Formula (18). By replacing the quantized linear prediction coefficient inverse adjustment unit 155 and the inverseadjusted LSP generating unit 160 in the first embodiment with the LSP linear transformation unit 300, a series of approximations of the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] can be generated with a smaller amount of calculation than the first embodiment.
The encoding apparatus 3 in the second embodiment decides whether to code in the time domain or in the frequency domain based on the magnitude of temporal variation in the input sound signal for each frame. However, even for a frame in which the temporal variation in the input sound signal was large and frequency domain encoding was selected, it is possible that actually a sound signal reproduced by encoding in the time domain leads to smaller distortion relative to the input sound signal than a signal reproduced by encoding in the frequency domain. Likewise, even for a frame in which the temporal variation in the input sound signal was small and encoding in the time domain was selected, it is possible that actually a sound signal reproduced by encoding in the frequency domain leads to smaller distortion relative to the input sound signal than a sound signal reproduced by encoding in the time domain. That is to say, the encoding apparatus 3 in the second embodiment cannot always select one of the time domain and frequency domain encoding methods that provides smaller distortion relative to the input sound signal. To address this, a encoding apparatus 8 in a modification of the second embodiment performs both time domain and frequency domain encoding on each frame and selects either of them that yields smaller distortion relative to the input sound signal.
<Encoding Apparatus>
The encoding apparatus 8 differs from the encoding apparatus 3 in the second embodiment in that it does not include the feature amount extracting unit 120 and includes a code selection and output unit 375 in place of the output unit 175.
<Encoding Method>
Referring to
In the encoding method according to the modification of the second embodiment, the LSP generating unit 110, LSP encoding unit 115, linear prediction coefficient adjusting unit 125, adjusted LSP generating unit 130, adjusted LSP encoding unit 135, quantized linear prediction coefficient generating unit 140, first quantized smoothed power spectral envelope series calculating unit 145, delay input unit 165, and LSP linear transformation unit 300 are also executed in addition to the input unit 100 and the linear prediction analysis unit 105 for all frames regardless of whether the temporal variation in the input sound signal is large or small. The operations of these components are the same as the second embodiment. However, the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }generated by the LSP linear transformation unit 300 is input to the delay input unit 165.
The delay input unit 165 holds the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] input from the LSP encoding unit 115 and the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }input from the LSP linear transformation unit 300 at least for the duration of one frame. When the frequency domain encoding method was selected by the code selection and output unit 375 for the preceding frame (i.e., when the identification code Cg output by the code selection and output unit 375 for the preceding frame is information indicating the frequency domain encoding method), the delay input unit 165 outputs the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }for the preceding frame input from the LSP linear transformation unit 300 to the time domain encoding unit 170 as the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] for the preceding frame. When the time domain encoding method was selected by the code selection and output unit 375 for the preceding frame (i.e., when the identification code Cg output by the code selection and output unit 375 for the preceding frame is information indicating the time domain encoding method), the delay input unit 165 outputs the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] for the preceding frame input from the LSP encoding unit 115 to the time domain encoding unit 170 (step S165).
As with the frequency domain encoding unit 150 in the second embodiment, the frequency domain encoding unit 150 generates and outputs frequency domain signal codes, and also determines and outputs the distortion or an estimated value of the distortion of the sound signal corresponding to the frequency domain signal codes relative to the input sound signal. The distortion or an estimation thereof may be determined either in the time domain or in the frequency domain. This means that the frequency domain encoding unit 150 may determine the distortion or an estimated value of the distortion of a frequencydomain sound signal series corresponding to frequency domain signal codes relative to the frequencydomain sound signal series that is obtained by converting the input sound to signal into the frequency domain.
The time domain encoding unit 170, as with the time domain encoding unit 170 in the second embodiment, generates and outputs time domain signal codes, and also determines the distortion or an estimated value of the distortion of the sound signal corresponding to the time domain signal codes relative to the input sound signal.
Input to the code selection and output unit 375 are the frequency domain signal codes generated by the frequency domain encoding unit 150, the distortion or an estimated value of distortion determined by the frequency domain encoding unit 150, the time domain signal codes generated by the time domain encoding unit 170, and the distortion or an estimated value of distortion determined by the time domain encoding unit 170.
When the distortion or estimated value of distortion input from the frequency domain encoding unit 150 is smaller than the distortion or an estimated value of distortion input from the time domain encoding unit 170, the code selection and output unit 375 outputs the frequency domain signal codes and identification code Cg which is information indicating the frequency domain encoding method. When the distortion or estimated value of distortion input from the frequency domain encoding unit 150 is greater than the distortion or an estimated value of distortion input from the time domain encoding unit 170, the code selection and output unit 375 outputs the time domain signal codes and identification code Cg which is information indicating the time domain encoding method. When the distortion or an estimated value of distortion input from the frequency domain encoding unit 150 is equal to the distortion or an estimated value of distortion input from the time domain encoding unit 170, the code selection and output unit 375 outputs either the time domain signal codes or the frequency domain signal codes according to predetermined rules, as well as identification code Cg which is information indicating the encoding method corresponding to the codes being output. That is to say, of the frequency domain signal codes input from the frequency domain encoding unit 150 and the time domain signal codes input from the time domain encoding unit 170, the code selection and output unit 375 outputs either one that leads to a smaller distortion of the sound signal reproduced from the codes relative to the input sound signal, and also outputs information indicative of the encoding method that yields smaller distortion as identification code Cg (step S375).
The code selection and output unit 375 may also be configured to select either one of the sound signals reproduced from the respective codes that has smaller distortion relative to the input sound signal. In such a configuration, the frequency domain encoding unit 150 and the time domain encoding unit 170 reproduce sound signals from the codes and output them instead of distortion or an estimated value of distortion. The code selection and output unit 375 outputs either the sound signal reproduced by the frequency domain encoding unit 150 or the sound signal reproduced by the time domain encoding unit 170 respectively from frequency domain signal codes and time domain signal codes that has smaller distortion relative to the input sound signal, and also outputs information indicating the encoding method that yields smaller distortion as identification code Cg.
Alternatively, the code selection and output unit 375 may be configured to select either one that has a smaller code amount. In such a configuration, the frequency domain encoding unit 150 outputs frequency domain signal codes as in the second embodiment. The time domain encoding unit 170 outputs time domain signal codes as in the second embodiment. The code selection and output unit 375 outputs either the frequency domain signal codes or the time domain signal codes that have a smaller code amount, and also outputs information indicating the encoding method that yields a smaller code amount as identification code Cg.
<Decoding Apparatus>
A code sequence output by the encoding apparatus 8 in the modification of the second embodiment can be decoded by the decoding apparatus 4 of the second embodiment as with a code sequence output by the encoding apparatus 3 of the second embodiment.
<Effects of Modification of the Second Embodiment>
The encoding apparatus 8 in the modification of the second embodiment provides similar effects to the encoding apparatus 3 of the second embodiment and further has the effect of reducing the code amount to be output compared to the encoding apparatus 3 of the second embodiment.
The encoding apparatus 1 of the first embodiment and the encoding apparatus 3 of the second embodiment once convert the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . . {circumflex over ( )}θ_{γR}[p] into linear prediction coefficients and then calculate the quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N]. A encoding apparatus 5 in the third embodiment directly calculates the quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] from the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . . {circumflex over ( )}θ_{γR}[p] without converting the adjusted quantized LSP parameter sequence to linear prediction coefficients. Similarly, a decoding apparatus 6 in the third embodiment directly calculates the decoded smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] from the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . . {circumflex over ( )}θ_{γR}[p] without converting the decoded adjusted LSP parameter sequence to linear prediction coefficients.
<Encoding Apparatus>
The encoding apparatus 5 differs from the encoding apparatus 3 in the second embodiment in that it does not include the quantized linear prediction coefficient generating unit 140 and the first quantized smoothed power spectral envelope series calculating unit 145 but includes a second quantized smoothed power spectral envelope series calculating unit 146 instead.
<Encoding Method>
Referring to
At step S146, the second quantized smoothed power spectral envelope series calculating unit 146 uses the adjusted quantized LSP parameters {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . . {circumflex over ( )}θ_{γR}[p] output by the adjusted LSP encoding unit 135 to determine a quantized smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] according to Formula (19) and outputs it.
<Decoding Apparatus>
The decoding apparatus 6 differs from the decoding apparatus 4 in the second embodiment in that it does not include the decoded linear prediction coefficient generating unit 220 and the first decoded smoothed power spectral envelope series calculating unit 225 but includes a second decoded smoothed power spectral envelope series calculating unit 226 instead.
<Decoding Method>
Referring to
At step S226, as with the second quantized smoothed power spectral envelope series calculating unit 146, the second decoded smoothed power spectral envelope series calculating unit 226 uses the decoded adjusted LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . . {circumflex over ( )}θ_{γR}[p] to determine a decoded smoothed power spectral envelope series {circumflex over ( )}W_{γR}[1], {circumflex over ( )}W_{γR}[2], . . . , {circumflex over ( )}W_{γR}[N] according to the Formula (19) above and outputs it.
The quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p] is a series that satisfies
0<{circumflex over ( )}θ[1]< . . . <{circumflex over ( )}θ[p]<π.
That is, it is a series in which parameters are arranged in ascending order. Meanwhile, the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }generated by the LSP linear transformation unit 300 is produced through approximate transformation, so it could not be in ascending order. To address this, the fourth embodiment adds processing for rearranging the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }output by the LSP linear transformation unit 300 into ascending order.
<Encoding Apparatus>
The encoding apparatus 7 differs from the encoding apparatus 5 in the second embodiment in that it further includes an approximate LSP series modifying unit 700.
<Encoding Method>
Referring to
The approximate LSP series modifying unit 700 outputs a series in which the values {circumflex over ( )}θ[i]_{app }in the approximate quantized LSP parameter sequence {circumflex over ( )}θ[1]_{app}, {circumflex over ( )}θ[2]_{app}, . . . , {circumflex over ( )}θ[p]_{app }output by the LSP linear transformation unit 300 have been rearranged in ascending order as a modified approximate quantized LSP parameter sequence {circumflex over ( )}θ′[1]_{app}, {circumflex over ( )}θ′[2]_{app}, . . . , {circumflex over ( )}θ′[p]_{app}. The modified first approximate quantized LSP parameter sequence {circumflex over ( )}θ′[1]_{app}, {circumflex over ( )}θ′[2]_{app}, . . . , {circumflex over ( )}θ′[p]_{app }output by the approximate LSP series modifying unit 700 is input to the delay input unit 165 as the quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}[2], . . . , {circumflex over ( )}θ[p].
In addition to merely rearranging the values in the approximate quantized LSP parameter sequence, each value {circumflex over ( )}θ[i]_{app }may be adjusted as {circumflex over ( )}θ′[i]_{app }such that {circumflex over ( )}θ[i+1]_{app}−{circumflex over ( )}θ[i]_{app} is equal to or greater than a predetermined threshold for each value of i=1, . . . , p−1.
While the foregoing embodiments were described assuming use of LSP parameters, an ISP parameter sequence may be employed instead of an LSP parameter sequence. An ISP parameter sequence ISP[1], . . . , ISP[p] is equivalent to a series consisting of an LSP parameter sequence of the p−1th order and PARCOR coefficient k_{p }of the pth order (the highest order). That is to say,
ISP[i]=θ[i] for i=1, . . . , p−1, and
ISP[p]=k_{p}.
Specific processing will be illustrated for a case where input to the LSP linear transformation unit 300 is an ISP parameter sequence in the second embodiment.
Assume that input to the LSP linear transformation unit 300 is a adjusted quantized ISP parameter sequence {circumflex over ( )}ISP_{γR}[1], {circumflex over ( )}ISP_{γR}[2], . . . , {circumflex over ( )}ISP_{γR}[p]. Here,
{circumflex over ( )}ISP_{γR}[1]={circumflex over ( )}θ_{γR}[i], and
{circumflex over ( )}ISP_{γR}[p]={circumflex over ( )}k_{p}.
The value {circumflex over ( )}k_{p }is the quantized value of k_{p}.
The LSP linear transformation unit 300 determines an approximate quantized ISP parameter sequence {circumflex over ( )}ISP[1]_{app}, . . . , {circumflex over ( )}ISP[p]_{app }through the following process and outputs it.
(Step 1) Given {circumflex over ( )}Θ_{γ1}=({circumflex over ( )}ISP_{γR}[1], . . . , {circumflex over ( )}ISP_{γR}[p−1])^{T}, p is replaced with p−1, and {circumflex over ( )}θ[1]_{app}, . . . , {circumflex over ( )}θ[p−1]_{app }are determined by calculating Formula (18). Here,
{circumflex over ( )}ISP[i]_{app=}{circumflex over ( )}θ[i]_{app }(i=1, . . . , p−1).
(Step 2) {circumflex over ( )}ISP[p]_{app }defined by the formula below is determined.
{circumflex over ( )}ISP[p]_{app}={circumflex over ( )}ISP_{γR}[p]·(1/γR)^{p}.
The LSP linear transformation unit 300 included in the encoding apparatuses 3, 5, 7, 8 and the decoded LSP linear transformation unit 400 included in the decoding apparatuses 4, 6 may also be implemented as a separate frequency domain parameter sequence generating apparatus.
The following description illustrates a case where the LSP linear transformation unit 300 included in the encoding apparatuses 3, 5, 7, 8 and the decoded LSP linear transformation unit 400 included in the decoding apparatuses 4, 6 are implemented as a separate frequency domain parameter sequence generating apparatus.
<Frequency Domain Parameter Sequence Generating Apparatus>
A frequency domain parameter sequence generating apparatus 10 according to the fifth embodiment includes a parameter sequence converting unit 20 for example, as shown in
The frequency domain parameters ω[1], ω[2], . . . , ω[p] to be input are a frequency domain parameter sequence derived from linear prediction coefficients, a[1], a[2], . . . , a[p], which are obtained by linear prediction analysis of sound signals in a predetermined time segment. The frequency domain parameters ω[1], ω[2], . . . , ω[p] may be an LSP parameter sequence θ[1], θ[2], . . . , θ[p] used in conventional encoding methods, or a quantized LSP parameter sequence {circumflex over ( )}θ[1], {circumflex over ( )}θ[2], . . . , {circumflex over ( )}θ[p], for example. Alternatively, they may be the adjusted LSP parameter sequence θ_{γR}[1], θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] or the adjusted quantized LSP parameter sequence {circumflex over ( )}θ_{γR}[1], {circumflex over ( )}θ_{γR}[2], . . . , {circumflex over ( )}θ_{γR}[p] used in the aforementioned embodiments, for example. Further, they may be frequency domain parameters equivalent to LSP parameters, such as the ISP parameter sequence described in the modification above, for example. A frequency domain parameter sequence derived from linear prediction coefficients a[1], a[2], . . . , a[p] are a series in the frequency domain derived from a linear prediction coefficient sequence and represented by the same number of elements as the order of prediction, typified by an LSP parameter sequence, an ISP parameter sequence, an LSF parameter sequence, or an ISF parameter sequence each derived from the linear prediction coefficient sequence a[1], a[2], . . . , a[p], or a frequency domain parameter sequence in which all of the frequency domain parameters ω[1], ω[2], . . . , ω[p−1] are present from 0 to π and, when all of the linear prediction coefficients contained in the linear prediction coefficient sequence are 0, the frequency domain parameters ω[1], ω[2], . . . , ω[p−1] are present from 0 to π at equal intervals.
The parameter sequence converting unit 20, similarly to the LSP linear transformation unit 300 and the decoded LSP linear transformation unit 400, applies approximate linear transformation to the frequency domain parameter sequence ω[1], ω[2], . . . , ω[p−1] making use of the nature of LSP parameters to generate a converted frequency domain parameter sequence ˜ω[1], ˜ω[2], . . . , ˜ω[p]. The parameter sequence converting unit 20 determines the value of the converted frequency domain parameter ˜ω[i] according to one of the methods shown below for each i=1, 2, . . . , p, for example.
1. The value of the converted frequency domain parameter ˜ω[i] is determined by linear transformation which is based on the relationship of values between ω[i] and one or more frequency domain parameters adjacent to ω[i]. For instance, linear transformation is performed so that the intervals between parameter values becomes more uniform or less uniform in the converted frequency domain parameter sequence ˜ω[i] than in the frequency domain parameter sequence ω[i]. Linear transformation that makes the parameter interval more uniform corresponds to processing that flats the waves of the amplitude of the power spectral envelope in the frequency domain (processing for smoothing the power spectral envelope). Linear transformation that makes the parameter interval less uniform corresponds to processing that emphasizes the height difference in the waves of the amplitude of the power spectral envelope in the frequency domain (processing for unsmoothing the power spectral envelope).
2. When ω[i] is closer to ω[i+1] relative to the midpoint between ω[i+1] and ω[i−1], then ˜ω[i] is determined so that ˜ω[i] will be closer to ˜ω[i+1] relative to the midpoint between ˜ω[i+1] and ˜ω[i−1] and that the value of ˜ω[i+1]−˜ω[i] will be smaller than ω[i+1]−ω[i]. When ω[i] is closer to ω[i−1] relative to the midpoint between ω[i+1] and ω[i−1], then ˜ω[i] is determined so that ˜ω[i] will be closer to ˜ω[i−1] relative to the midpoint between ˜ω[i+1] and ˜ω[i−1] and that the value of ˜ω[i]−˜ω[i−1] will be smaller than ω[i]−ω[i−1]. This corresponds to processing that emphasizes the height difference in the waves of the amplitude of the power spectral envelope in the frequency domain (processing for unsmoothing the power spectral envelope).
3. When ω[i] is closer to ω[i+1] relative to the midpoint between ω[i+1] and ω[i−1], then ˜ω[i] is determined so that ˜ω[i] will be closer to ˜ω[i+1] relative to the midpoint between ˜ω[i+1] and ˜ω[i−1] and that the value of ˜ω[i+1]−˜ω[i] will be greater than ω[i+1]−ω[i]. When ω[i] is closer to ω[i−1] relative to the midpoint between ω[i+1] and ω[i−1], then ˜ω[i] is determined so that ˜ω[i] will be closer to ˜ω[i−1] relative to the midpoint between ˜ω[i+1] and ˜ω[i−1] and that the value of ˜ω[i]−˜ω[i−1] will be greater than ω[i]−ω[i−1]. This corresponds to processing that flats the waves of the amplitude of the power spectral envelope in the frequency domain (processing for smoothing the power spectral envelope).
For example, the parameter sequence converting unit 20 determines the converted frequency domain parameters ˜ω[1], ˜ω[2], . . . , ˜ω[p] according to Formula (20) below and outputs it.
Here, γ1 and γ2 are positive coefficients equal to or smaller than 1. Formula (20) can be derived by setting Θ_{γ1}=(ω[1], ω[2], . . . , ω[p])^{T }and Θ_{γ2}=(˜ω[1], ˜ω[2], . . . , ˜ω[p])^{T }in Formula (13), which models LSP parameters, and defining
In this case, frequency domain parameters ω[1], ω[2], . . . , ω[p] are a frequencydomain parameter sequence or the quantized values thereof equivalent to
a[1]×(γ1), a[2]×(γ1)^{2}, . . . , a[p]×(γ1)^{p},
which is a coefficient sequence that has been adjusted by multiplying each coefficient a[i] of the linear prediction coefficients a[1], a[2], . . . , a[p] by the ith power of the factor γ1. The converted frequency domain parameters ˜ω[1], ˜ω[2], . . . , ˜ω[p] are a series that approximates a frequencydomain parameter sequence equivalent to
a[1]×(γ2), a[2]×(γ2)^{2}, . . . , a[p]×(γ2)^{p},
which is a coefficient sequence that has been adjusted by multiplying each coefficient a[i] of the linear prediction coefficients a[1], a[2], . . . , a[p] by the ith power of factor γ2.
<Effects of the Fifth Embodiment>
As with the encoding apparatuses 3, 5, 7, 8 or the decoding apparatuses 4, 6, the frequency domain parameter sequence generating apparatus in the fifth embodiment is able to determine converted frequency domain parameters from frequency domain parameters with a smaller amount of calculation than when converted frequency domain parameters are determined from frequency domain parameters by way of linear prediction coefficients as in the encoding apparatus 1 and the decoding apparatus 2.
The present invention is not limited to the abovedescribed embodiments and it goes without saying that modifications may be made as necessary without departing from the scope of the invention. The various kinds of processing illustrated in the embodiments above could also be performed in parallel or separately in accordance with the processing capability of the device executing them or certain necessity in addition to being carried out chronologically in the orders described herein.
[Program and Recording Media]
When the various processing functions of the apparatuses described in the embodiments are implemented by a computer, the processing details of the functions supposed to be provided in the apparatuses are described by a program. The program is then executed by the computer so as to implement various processing functions of the individual apparatuses on the computer.
A program describing the processing details can be recorded in a computerreadable recording medium. The computerreadable recording medium may be any kind of media, such as a magnetic recording device, optical disk, magnetooptical recording medium, and semiconductor memory, for example.
Such a program may be distributed by selling, granting, or lending a portable recording medium, such as a DVD or CDROM for example, having the program recorded thereon. Alternatively, the program may be stored in a storage device at a server computer and transferred to other computers from the server computer over a network so as to distribute the program
When a computer is to execute such a program, the computer first stores the program recorded on a portable recording medium or the program transferred from the server computer once in its own storage device, for example. Then, when it carries out processing, the computer reads the program stored in its recording medium and performs processing in accordance with the program that has been read. As an alternative form of execution of the program, the computer may directly read the program from a portable recording medium and perform processing in accordance with the program, or the computer may perform processing sequentially in accordance with a program it has received every time a program is transferred from the server computer to the computer. The abovedescribed processing may also be implemented as a socalled application service provider (ASP) service, which implements processing functions only through requests for execution and acquisition of results without transfer of programs from a server computer to a computer. Programs in the embodiments described herein are intended to contain information that is used in processing by an electronic computer and subordinate to programs (such as data that is not a direct instruction on a computer but has properties governing the processing of the computer).
Additionally, while the apparatuses of the present invention have been described as being implemented through execution of predetermined programs on computer in such embodiments, at least part of these processing details may also be implemented by hardware.