Pattern and speech recognition using accumulated partial scores from a posteriori odds, with pruning based on calculation amount

• US 5,749,069 A
• Filed: 03/17/1995
• Issued: 05/05/1998
• Est. Priority Date: 03/18/1994
• Status: Expired due to Fees

• Current Assignee: ATR Human Information Processing Research Laboratories
• Original Assignee: ATR Human Information Processing Research Laboratories
• Inventors: Komori, Takashi, Katagiri, Shigeru
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Smits, Talivaldis Ivars
• Application Number:
  US08/405,674
• Time in Patent Office: 1,145 Days
• Field of Search: 395/20, 395/2.42, 395/2.49, 395/2.6, 395/2.61, 395/2.63, 395/2.64
• US Class Current: 704/240
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

A speech recognition apparatus includes a data input portion to which input data which is a speech pattern is applied, a score calculation portion calculating a score indicating a possibility of recognition of a partial pattern of the speech pattern based on the estimate of a posteriori odds, an optimization design portion designing optimized parameters for calculating the estimate of the a posteriori odds in the score calculation portion and/or optimized parameters of pruning functions controlling calculation amount in the pruning processing portion, a pruning processing portion pruning the score for making operation efficient, an accumulated score calculating portion accumulating pruned scores to calculate an accumulated score, a recognition result decision portion classifying input data for every class based on the accumulated score and deciding a recognition result, and a recognition result output portion providing the recognition result.

Claims

1. A pattern recognition method for classifying a continuous pattern representative of a physical activity or phenomenon to one or predetermined classes each indicating a combination of one or more predetermined subclasses, comprising:

a first step of determining a plurality of partial scores each indicating an estimate of logarithm of a posteriori odds that one of partial patterns of said continuous pattern is classified to one of said subclasses;

a second step of determining accumulated scores each indicating summation of one or more of said partial scores; and

a third step of classifying said continuous pattern into one of said predetermined classes based on said determined accumulated scores, wherein

scores of any possible partial patterns are used, and

the partial patterns are not first isolated.

2. The pattern recognition method as recited in claim 1, wherein

said first step includes the step of pruning candidates of said partial patterns for determination of said accumulated score or classification of said continuous pattern.

3. The pattern recognition method as recited in claim 2, wherein

said first step includes the step of finding the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimated a posteriori odds for determination of said score.

4. The pattern recognition method as recited in claim 2, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

5. The pattern recognition method as recited in claim 2, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

6. The pattern recognition method as recited in claim 2, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of said partial patterns.

7. The pattern recognition method as recited in claim 2, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of said partial patterns.

8. The pattern recognition method as recited in claim 1, wherein

said first step includes the step of finding the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimated a posteriori odds for determination of said score.

9. The pattern recognition method as recited in claim 1, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

10. The pattern recognition method as recited in claim 1, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

11. A speech recognition method for classifying a time series speech pattern representative of a physical activity or phenomenon to one of predetermined classes each indicating a combination of one or more predetermined subclasses, comprising:

a first step of determining a plurality of partial scores each indicating an estimate of logarithm of a posteriori odds that one of partial patterns of said speech pattern is classified to one of said subclasses;

a second step of determining accumulated scores each indicating summation of one or more of said partial scores; and

a third step of classifying said speech pattern into one of said predetermined classes based on said determined accumulated scores, wherein

scores of any possible partial patterns are used, and

the partial patterns are not first isolated.

12. The speech recognition method as recited in claim 11, wherein

said first step includes the step of pruning candidates of said partial patterns for determination of said accumulated score or classification of said speech pattern.

13. The speech recognition method as recited in claim 12, wherein

said first step includes the step of finding the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimated a posteriori odds for determination of said score.

14. The speech recognition method as recited in claim 12, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

15. The speech recognition method as recited in claim 12, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

16. The speech recognition method as recited in claim 12, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of candidates of said partial patterns.

17. The speech recognition method as recited in claim 12, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of candidates of said partial patterns.

18. The speech recognition method as recited in claim 11, wherein

said first step includes the step of finding the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimate of the a posteriori odds for determination of said score.

19. The speech recognition method as recited in claim 11, wherein

said first step includes the step of representing a classification error rate as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

20. The speech recognition method as recited in claim 11, wherein

said first step includes the step of representing a classification error rate and a calculation amount as a predetermined function and finding the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

21. A speech recognition apparatus for classifying a time series speech pattern representative of a physical activity or phenomenon to one of predetermined classes each indicating a combination of one or more predetermined subclasses, comprising:

score determination means for determining a plurality of partial scores each indicating an estimate of logarithm of a posteriori odds that one of partial patterns of said speech pattern is classified to one of said subclasses;

accumulated score determination means for determining accumulated scores each indicating summation of one or more of said partial scores; and

classification means for classifying said speech pattern into one of said predetermined classes based on said determined accumulated scores, wherein

scores of any possible partial patterns are used, and

the partial patterns are not first isolated.

22. The speech recognition apparatus as recited in claim 21, further comprising:

pruning means for pruning candidates of said partial patterns for determination of said accumulated score or classification of said speech pattern.

23. The speech recognition apparatus as recited in claim 22, further comprising:

optimization means for optimizing the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimated a posteriori odds for determination of said score.

24. The speech recognition apparatus as recited in claim 22, further comprising:

optimization means for representing a classification error rate as a predetermined function and optimizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

25. The speech recognition apparatus as recited in claim 22, further comprising:

optimization means for representing a classification error rate and a calculation amount as a predetermined function and optimizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

26. The speech recognition apparatus as recited in claim 22, further comprising:

optimization means for representing a classification error rate as a predetermined function and optimizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of candidates of said partial patterns.

27. The speech recognition apparatus as recited in claim 22, further comprising:

optimization means for representing a classification error rate and a calculation amount as a predetermined function and optimizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score and pruning of candidates of said partial patterns.

28. The speech recognition apparatus as recited in claim 21, further comprising:

optimization means for optimizing the estimate of said a posteriori odds using a parameter maximizing likelihood of the estimated a posteriori odds for determination of said score.

29. The speech recognition apparatus as recited in claim 21, further comprising:

optimization means for representing a classification error rate as a predetermined function and optimizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.

30. The speech recognition apparatus as recited in claim 21, further comprising:

optimization means for representing a classification error rate and a calculation amount as a predetermined function and maximizing the estimate of said a posteriori odds using a parameter minimizing the value of the function for determination of said score.