Normalization of speech signals

• US 4,817,158 A
• Filed: 10/19/1984
• Issued: 03/28/1989
• Est. Priority Date: 10/19/1984
• Status: Expired due to Term

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Picheny, Michael A.
• Primary Examiner(s): Kemeny, Emanuel S.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US06/662,867
• Time in Patent Office: 1,621 Days
• Field of Search: 381/36-43, 381/46-52, 381/94, 382/18
• US Class Current: 704/224
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit

Abstract

A method and a system are disclosed for normalizing a speech signal prior to a speech recognition process. In a preparatory procedure, a sample interval of speech is separated into thirty-one frequency bands and an amplitude histogram is generated for each band. From these histograms, the 5% percentile amplitude value P(05) and the 95% percentile amplitude value P(95) are extracted for each band and these values are stored for later reference. For actual normalization, the current speech signal is also divided into the same frequency band as in the preparatory procedure, and consecutive input amplitude values A(in) of each frequency band are modified, using the percentile values of the respective band, to obtain output values according to ##EQU1## The essential effect of this normalizing treatment is that the resulting long-term spectrum is given by the P(95) values and the spectrum of silence is given by the P(05) values. After normalization, all speech has the same silence spectrum and long-term spectrum.

Claims

1. A method for overcoming the distortions in the spectrum-of-silence in a system for accepting words presented in a stream of continuous speech, processing the stream into amplitude histograms for respective frequencies, and carrying out recognition processes, characterized by

(a) collecting amplitude histograms as a function of frequency;

(b) extracting, for each frequency, the amplitude P(05) defining the 5th percentile and the amplitude P(95) defining the 95th percentile; and

(c) normalizing, for each frequency, the input amplitude A(in) of the speech signal to obtain an output signal amplitude ##EQU6##

2. System for normalizing a speech signal for a subsequent analysis or recognition process, characterized by

means for generating an amplitude histogram for each of several frequency bands, comprising for each of k consecutive amplitude ranges in each frequency band, a histogram count of speech signal amplitudes falling into the respective amplitude range;

means for extracting, for each frequency from its histogram, two percentile amplitude values P(05) and P(95) indicative of the 5th and 95th percentile of the histogram counts, respectively.

means for storing said percentile amplitude values; and

means responsive to consecutive amplitude values A(in) of a current speech signal and to the percentile amplitude values available in said storing means, for generating normalized speech signal amplitude values ##EQU7## for each said frequency band.

3. A method for normalization a speech signal for subsequent speech recognition, the method comprising the steps of:

(a) filtering an input speech signal into a plurality of frequency bands;

(b) sampling the input speech signal at a prescribed sampling rate;

(c) for one sample after another in an analysis interval of the input speech signal, detecting amplitudes values of the input speech signal in each frequency band;

(d) generating, based on the amplitude values detected in step (c), a respective low percentile amplitude value P(lo) and a respective high percentile amplitude value P(hi) for each frequency band and storing the generated percentile amplitude values; and

(e) for an amplitude value A(in) of the input speech signal in a given frequency band f, adapting the value A(in) to;

##EQU8##

4. Method according to claim 3, wherein the percentile generating step includes the steps of:

forming a plurality of histograms each of which indicates the number of detected amplitudes corresponding to a given amplitude range

(i) for a given frequency band, and

(ii) for a given analysis interval; and

determining percentile amplitude values from the formed histograms.

5. Method according to claim 3, characterized in that

the high percentile amplitude value for a given subband corresponds to the 95% percentile of all amplitude values of the given subband, and the low percentile amplitude value of the given subband corresponds to the 5% percentile of all amplitude values of the given subband.

6. Method according to claim 4, characterized in that

the analysis internal for initially obtaining the high and low percentile amplitude values is one minute.

7. Method according to claim 3, characterized in that

the input speech signal is filtered, for the generation of percentile amplitude values and for A(in) value adapting, into more than 30 frequency bands.

8. Method according to claim 7, characterized in that

the speech signal is filtered into eight lower frequency bands of equal bandwidth between 200 and 1,000 Hertz, and that the remainder of the spectrum from 1,000 Hertz up to the maximum frequency of the prefiltered speech signal is filtered into bands according to a logarithmic schedule.

9. The method of claim 4 comprising the further step of:

digitizing the input speech signal wherein said sampling and digitizing steps are prior to the filtering step (a).

10. The method of claim 4 wherein the sampling rate is on the order of 1 sample per 0.01 seconds and wherein at least some of the amplitude ranges are on the order of 1 dB.