Speech recognition arrangement

• US 4,937,870 A
• Filed: 11/14/1988
• Issued: 06/26/1990
• Est. Priority Date: 11/14/1988
• Status: Expired due to Term

• Current Assignee: American Telephone & Telegraph Company (AT&T, Inc.), Bell Telephone Laboratories, Inc. (Nokia Corporation)
• Original Assignee: American Telephone & Telegraph Company (AT&T, Inc.)
• Inventors: Bossemeyer, Robert W. Jr.
• Primary Examiner(s): Harkcom, Gary V.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US07/269,527
• Time in Patent Office: 589 Days
• Field of Search: 364/513.5, 381/41-43
• US Class Current: 704/241
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]

Abstract

A speech recognition arrangement where nonspectral features of the input signal, e.g., energy and voicing parameters, are used to effectively remove non-speech events from consideration but only after a time warping procedure based solely on the input and reference pattern spectral parameters has been completed. The time warping procedure is not unduly complex because there is no need to weight spectral and nonspectral parameters in matching input and reference patterns. For each reference pattern, the time warping procedure defines a scan region of the input signal to be used in evaluating the nonspectral input signal characteristics. Energy and voicing parameters are useful in distinguishing non-speech events since speech patterns typically have few very low-energy frames (other than frames that are part of a gap within a vocabulary item) and more than a minimum number of voiced frames, e.g., frames corresponding to vowel sounds.

Claims

1. In a speech recognizer having a plurality of stored reference pattern templates each comprising a time frame sequence of acoustic spectral parameters of a prescribed reference pattern, a method for processing an input signal to recognize a speech pattern comprising

generating a time frame sequence of acoustic spectral parameters from said input signal,

generating a time frame sequence of acoustic nonspectral parameters from said input signal,

time aligning each of said reference pattern templates with said input signal based on reference pattern and input signal spectral parameters but independent of said nonspectral parameters,

determining a set of similarity measures each representative of the similarity between spectral parameters of said input signal and spectral parameters of one of the time aligned reference pattern templates and

selectively identifying said speech pattern in said input signal as one of said reference patterns based both on said similarity measures and on said nonspectral parameters,

wherein said time aligning comprises

for each of said reference patterns, pairing time frames of that reference pattern template with time frames of said input signal to maximize the similarity measure determined for that reference pattern, said pairing defining a scan region of input signal time frames for that reference pattern,

wherein said selectively identifying comprises

for each of said reference patterns, adjusting the determined similarity measured based on said nonspectral parameters over the scan region of input signal time frames for that reference pattern and

selectively identifying said speech pattern in said input signal as one of said reference patterns based on said adjusted similarity measures.

2. A method in accordance with claim 1 wherein said sequence of acoustic nonspectral parameters comprises energy parameters and wherein said adjusting comprises

for each of said reference patterns, adjusting the determined similarity measure based on the number of input signal time frames of the scan region for that reference pattern that have an energy parameter less than a predetermined threshold.

3. A method in accordance with claim 1 wherein said sequence of acoustic nonspectral parameters comprises energy parameters and wherein said adjusting comprises

for each of said reference patterns, adjusting the determined similarity measure based on the number of input signal time frames of the scan region for that reference pattern that have an energy parameter less than a predetermined threshold and that are not part of a gap within a vocabulary item.

4. A method in accordance with claim 1 wherein said sequence of acoustic nonspectral parameters comprises voicing parameters and wherein said adjusting comprises

for each of said reference patterns, adjusting the determined similarity measured based on the number of input signal time frames of the scan region for that reference pattern that have a voicing parameter defining the frame as voiced.

5. A method in accordance with claim 1 wherein said sequence of acoustic nonspectral parameters comprises energy parameters and voicing parameters and wherein said adjusting comprises

for each of said reference patterns, adjusting the determined similarity measure based on the number of input signal time frames of the scan region for that reference pattern that have an energy parameter less than a predetermined threshold and that are not part of a gap within a vocabulary item and based on the number of input signal time frames of the scan region for that reference pattern that have a voicing parameter defining the frame as voiced.

6. A method in accordance with claim 1 wherein said sequence of acoustic nonspectral parameters comprises voicing parameters and wherein said adjusting comprises

for each of said reference patterns, adjusting the determined similarity measure based on the difference between a minimum number and the number of input signal time frames of the scan region for that reference pattern that have a voicing parameter defining the frame as voiced.

7. A method in accordance with claim 6 wherein said minimum number is constant for all of said reference patterns.

8. A method in accordance with claim 6 wherein said minimum number is a variable dependent on an expected number of voiced frames for a reference pattern.

9. In a speech recognizer having a plurality of stored reference pattern templates each comprising a time frame sequence of acoustic spectral parameters of a prescribed reference pattern, a method for processing an input signal to recognize a speech pattern comprising

generating a time frame sequence of acoustic spectral parameters from said input signal,

generating a time frame sequence of voicing parameters from said input signal, each of said voicing parameters defining the presence or absence of a vowel sound

determining a set of similarity measures each representative of the similarity between spectral parameters of said input signal and spectral parameters of one of the reference pattern templates and

selectively identifying said speech pattern in said input signal as one of said reference patterns based both on said similarity measures and on said voicing parameters.

10. A speech recognizer for processing an input signal to recognize a speech pattern comprising

memory means for storing a plurality of reference pattern templates each comprising a time frame sequence of acoustic spectral parameters of a prescribed reference pattern and

digital signal processor means comprising

means responsive to said input signal for generating a time frame sequence of acoustic spectral parameters,

means responsive to said input signal for generating a time frame sequence of acoustic nonspectral parameters,

means for time aligning each of said reference pattern templates with said input signal based on reference pattern and input signal spectral parameters but independent of said nonspectral parameters,

means for determining a set of similarity measures each representative of the similarity between spectral parameters of said input signal and spectral parameters of one of the time aligned reference pattern templates and

means for selectively identifying said speech pattern in said input signal as one of said reference patterns based both on said similarity measures and on said nonspectral parameters,

wherein said time aligning means comprises

means for pairing, for each of said reference patterns, time frames of that reference pattern template with time frames of said input signal to maximize the similarity measure determined by said determining means for that reference pattern, said pairing defining a scan region of input signal time frames for that reference pattern,

wherein said selectively identifying means comprises

means for adjusting, for each of said reference patterns, the determined similarity measure based on said at least one nonspectral parameter and

means for selectively identifying said speech pattern in said input signal as one of said reference patterns based on said adjusted similarity measures.

11. A speech recognizer for processing an input signal to recognize a speech pattern comprising

memory means for storing a plurality of reference pattern templates each comprising a time frame sequence of acoustic spectral parameters of a prescribed reference pattern and

digital signal processor means comprising

means responsive to said input signal for generating a time frame sequence of acoustic spectral parameters,

means responsive to said input signal for generating a time frame sequence of voicing parameters, each of said voicing parameters defining the presence or absence of a vowel sound,

means for determining a set of similarity measures each representative of the similarity between spectral parameters of said input signal and spectral parameters of one of the reference pattern templates and

means for selectively identifying said speech pattern in said input signal as one of said reference patterns based both on said similarity measures and on said voicing parameters.