Reference pattern generating apparatus and method, and computer readable medium having reference pattern generating program embodied thereon

• US 6,708,151 B1
• Filed: 09/12/2000
• Issued: 03/16/2004
• Est. Priority Date: 09/13/1999
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Hanazawa, Toshiyuki
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US09/660,478
• Time in Patent Office: 1,281 Days
• Field of Search: 704/243, 704/244, 704/254
• US Class Current: 704/244
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

A reference pattern generating apparatus comprises an initial reference pattern generating unit for dividing a time series of feature vectors obtained from a voice signal into a sequence of small sections in such a manner that any two adjacent small sections overlap each other, and for averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate an initial reference pattern having a plurality of states respectively associated with the plurality of small sections. A reference pattern generating unit brings each of the plurality of states of the initial reference pattern into correspondence with part of the time series of feature vectors using pattern matching, and then averages the part of the time series of feature vectors brought into correspondence with each of the plurality of states of the initial reference pattern so as to update the initial reference pattern to generate a reference pattern.

Claims

1. A reference pattern generating apparatus, comprising:

initial reference pattern generating means for dividing a time series of feature vectors obtained from a voice signal into a sequence of small sections in such a manner that any two adjacent small sections overlap each other, and for averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate an initial reference pattern having a plurality of states respectively associated with the plurality of small sections; and

reference pattern generating means for bringing each of the plurality of states of the initial reference pattern into correspondence with part of the time series of feature vectors using pattern matching, and for averaging the part of the time series of feature vectors brought into correspondence with each of the plurality of states of the initial reference pattern so as to update the initial reference pattern to generate a reference pattern.

2. The reference pattern generating apparatus according to claim 1, wherein said initial reference pattern generating means divides the time series of feature vectors into the sequence of small sections in such a manner that the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length.

3. The reference pattern generating apparatus according to claim 1, wherein said initial reference pattern generating means divides the time series of feature vectors into the sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion.

4. The reference pattern generating apparatus according to claim 1, wherein said initial reference pattern generating means selects a predetermined number of feature vectors exhibiting a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, the number of feature vectors selected being equal to a number of overlapped portions formed in the sequence of small sections, and then divides the time series of feature vectors into the sequence of small sections in such a manner that the plurality of overlapped portions include the predetermined number of feature vectors selected, respectively.

5. The reference pattern generating apparatus according to claim 1, wherein said initial reference pattern generating means includes:

a first initial value generating means for dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other and the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length, and for averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a first initial reference pattern;

a second initial value generating means for dividing the time series of feature vectors into a sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that any two adjacent small sections overlap each other, and one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion, and for averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a second initial reference pattern; and

third initial value generating means for selecting a predetermined number of feature vectors having a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, for dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other and the plurality of overlapped portions include the predetermined number of feature vectors selected, respectively, the predetermined number of feature vectors selected being equal to a number of overlapped portions formed in the sequence of small sections, and for averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a third initial reference pattern, and

wherein said reference pattern generating means generates three reference patterns from the first through third initial reference patterns generated by said first through third initial value generating means, and

said apparatus further comprises a reference pattern selecting means for performing pattern matching between each of the three reference patterns generated by said first through third initial value generating means and the time series of feature vectors and for selecting one reference pattern showing a minimum pattern matching distance from the three reference patterns.

6. A reference pattern generating method of compressing a time series of feature vectors obtained from a voice signal and generating a reference pattern, said method comprising the steps of:

dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other;

averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate an initial reference pattern having a plurality of states associated with the plurality of small sections; and

generating a reference pattern by bringing each of the plurality of states of the initial reference pattern into correspondence with part of the time series of feature vectors using pattern matching and by averaging the part of the time series of feature vectors brought into correspondence with each of the plurality of states of the initial reference pattern so as to update the initial reference pattern to generate a reference pattern.

7. The reference pattern generating method according to claim 6, wherein said dividing step includes dividing the time series of feature vectors into the sequence of small sections in such a manner that the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length.

8. The reference pattern generating method according to claim 6, wherein said dividing step includes dividing the time series of feature vectors into the sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion.

9. The reference pattern generating method according to claim 6, wherein said dividing step includes:

selecting a predetermined number of feature vectors having a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, the number of feature vectors selected being equal to a number of overlapped portions formed in the sequence of small sections, and

dividing the time series of feature vectors into the sequence of small sections in such a manner that the plurality of overlapped portions include the predetermined number of feature vectors selected, respectively.

10. The reference pattern generating method according to claim 6, wherein said dividing and averaging steps include:

dividing the time series of feature vectors into a sequence of small sections in such manner that any two adjacent small sections overlap each other, and the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a first initial reference pattern;

dividing the time series of feature vectors into a sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that any two adjacent small sections overlap each other and one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a second initial reference pattern; and

selecting a predetermined number of feature vectors having a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other and the plurality of overlapped portions include the predetermined number of feature vectors selected, respectively, the number of feature vectors selected being equal to a number of overlapped portions formed in the plurality of small sections, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a third initial reference pattern, and

wherein said reference pattern generating step includes generating three reference patterns;

said method further comprising performing pattern matching between each of the three reference patterns and the time series of feature vectors of and selecting one reference pattern showing a minimum pattern matching distance from the three reference patterns.

11. A computer readable medium having a computer program embodied thereon, the computer program being executable by a machine to the perform the steps of:

dividing the time series of feature vectors obtained from a voice signal into a sequence of small sections in such a manner that any two adjacent small sections overlap each other;

averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate an initial reference pattern having a plurality of states associated with the plurality of small sections; and

generating a reference pattern by bringing each of the plurality of states of the initial reference pattern into correspondence with part of the time series of feature vectors using pattern matching, and by averaging the part of the time series of feature vectors brought into correspondence with each of the plurality of states of the initial reference pattern so as to update the initial reference pattern.

12. The computer readable medium according to claim 11, wherein said dividing step includes dividing the time series of feature vectors into the sequence of small sections in such a manner that the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length.

13. The computer readable medium according to claim 11, wherein said dividing step includes dividing the time series of feature vectors into the sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion.

14. The computer readable medium according to claim 11, wherein said dividing step includes:

selecting a predetermined number of feature vectors having a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, the number of feature vectors selected being equal to a number of overlapped portions formed in the sequence of small sections, and dividing the time series of feature vectors into the sequence of small sections in such a manner that the plurality of overlapped portions include the predetermined number of feature vectors selected respectively.

15. The computer readable medium according to claim 11, wherein said dividing and averaging steps include:

dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other, and the plurality of small sections are equal in length where possible, otherwise all but one end small section are equal in length, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a first initial reference pattern;

dividing the time series of feature vectors into a sequence of small sections while making a distinction between pause portions and a speech portion included in the voice signal in such a manner that any two adjacent small sections overlap each other and one small section is assigned to each of the pause portions and remaining small sections are assigned to the speech portion, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a second initial reference pattern; and

selecting a predetermined number of feature vectors having a larger amount of change against an immediately preceding feature vector from the time series of feature vectors, dividing the time series of feature vectors into a sequence of small sections in such a manner that any two adjacent small sections overlap each other and the plurality of overlapped portions include the predetermined number of feature vectors selected, respectively, the number of feature vectors selected being equal to a number of overlapped portions formed in the plurality of small sections, and averaging part of the time series of feature vectors included in each of the plurality of small sections so as to generate a third initial reference pattern, and

wherein said reference pattern generating step is the step of generating three reference patterns from the first to third initial reference patterns and said computer program further performs the steps of performing pattern matching between each of the three reference patterns and the time series of feature vectors of and selecting one reference pattern showing a minimum pattern matching distance from the three reference patterns.