Speech processing apparatus and method employing matching and confidence scores

• US 7,072,836 B2
• Filed: 07/03/2001
• Issued: 07/04/2006
• Est. Priority Date: 07/12/2000
• Status: Expired due to Fees

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Shao, Yuan
• Primary Examiner(s): Young, W. R.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US09/898,776
• Publication Number:
  US 20020049593A1
• Time in Patent Office: 1,827 Days
• Field of Search: 704/251, 704/255, 704/252, 704/254
• US Class Current: 704/255
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence

Abstract

A speech recognition method and apparatus is disclosed in which each utterance matched to a feature model within a feature model memory (110) is associated with a confidence score indicative of the posterior probability of the word being correctly matched, given that the matching of the utterance to the feature model generated certain values indicative of the goodness of the match. The confidence score for the matching of an utterance to a feature model is determined from the generated values indicative of the goodness of the match and a stored set of parameters indicating the probability of the generated values arising given that a match is either correct or incorrect.

Claims

1. A speech processing apparatus comprising:

a data store operable to store a plurality of word models;

a receiver operable to receive an utterance; and

a matching unit operable to determine which of a plurality of word models stored in said data store most closely matches an utterance received by said receiver, said matching unit being arranged to output a value indicative of the goodness of each match between a most closely matching word model and an utterance received by said receiver;

characterized by further comprising;

a confidence model store operable to associate each word model stored in said data store with data indicative of the probability of a said value indicative of the goodness of match being output if said word model correctly or incorrectly matches an utterance; and

a confidence score output unit operable to determine and output a confidence score indicative of a calculated posterior probability that a received utterance has been correctly matched to a word model given that the match resulted in the output of a particular value indicative of goodness of match, said output unit being arranged to calculate said confidence score for a match utilizing said value output by said matching unit and said data associated with the word model matched to utterances by said confidence model store.

2. Apparatus in accordance with claim 1, wherein said matching unit is arranged to output a plurality of values indicative of the goodness of a match, wherein said confidence model store is arranged to associate data indicative of the probability of each of said output values being output, and wherein said output unit is arranged to calculate said confidence scores for matches utilizing said plurality of output values and said data associated with word models by said confidence model store.

3. Apparatus in accordance with claim 1, wherein said confidence model store is arranged to associate the same data with all word models stored within said data store.

4. An apparatus in accordance with claim 1, wherein said confidence model store is further arranged to associate with word models data indicative of the probability of word models being correctly or incorrectly matched to any utterance, wherein said output unit is arranged to utilize said data to calculate said confidence score.

5. Apparatus in accordance with claim 1, wherein said confidence model store is arranged to associate different data with different word models stored within said data store.

6. Apparatus in accordance with claim 5, wherein said confidence model store is arranged to associate groups of word models stored in said data store with different data.

7. Apparatus in accordance with claim 1, wherein said matching unit is arranged to determine match scores indicative of the correspondence between an utterance and word models, and output as a value indicative of the goodness of a match of the match score for the most closely matching word model.

8. Apparatus in accordance with claim 7, wherein said matching unit is arranged to output as a value indicative of the goodness of a match, data indicative of the ratio of a match score indicative of the correspondence between an utterance and a most closely matching word model and a match score for the correspondence between an utterance and a different word model.

9. Apparatus in accordance with claim 8, wherein said different word model comprises the second most closely matching word model to an utterance.

10. Apparatus in accordance with claim 8, wherein said different word model comprises the most closely matching word model to an utterance that is representative of a different word to said most closely matching word model.

11. Apparatus in accordance with claim 1, wherein said confidence model store is arranged to associate with each word model data defining a probability density function of the probabilities of a said value being output if said word model correctly or incorrectly matches an utterance.

12. An apparatus in accordance with claim 11, wherein said data defining a probability density function comprises data defining a function type and one or more parameters.

13. An apparatus in accordance with claim 11, wherein said output unit is arranged to determine probabilities that a value indicative of goodness of match would be output if a word is correctly or incorrectly matched utilizing said data defining probability functions associated with a matched word and to utilize said determined probabilities to calculate said confidence score.

14. A method of generating data for association with word models stored within a speech processing apparatus in accordance with claim 1, the method comprising the steps of:

determining for each utterance in a test vocabulary of known words, word models matched to said test utterances and said values output by said matching unit;

for utterances correctly matched to known words in said test vocabulary, determining a function corresponding to the probability of said values indicative of goodness of match arising;

for utterances incorrectly matched to known words in said test vocabulary, determining a function corresponding to the probability of values indicative of goodness of match arising; and

outputting data representative of said determined functions.

15. A method in accordance with claim 14, further comprising the steps of determining the proportion of utterances correctly or incorrectly matched to a word model and outputting data representative of the probability of a word model being correctly or incorrectly matched.

16. A method in accordance with claim 14, wherein said determination steps are performed for each word model stored in said speech processing apparatus and data is output for association with each of said models.

17. A method in accordance with claim 14, wherein said determination steps are performed for each word model stored in said speech processing apparatus representative of words of selected lengths and data is output for association with groups of models representative of words of said selected lengths.

18. A method in accordance with claim 14, further comprising the step of:

storing said output data in association with word models stored within a speech processing apparatus.

19. A method in accordance with claim 18, further comprising the steps of:

for a set of test utterances representative of known words, utilizing said speech processing apparatus to determine for said set of test utterances, a set of matched word models and output confidence scores;

for word models matched to test utterances, determining the difference between the sum of confidence scores for utterances matched to each said word model and the number of utterances correctly matched to said word model; and

generating further data for association with a word model if said difference is greater than a predetermined proportion of said test utterances matched to said word model.

20. A recording medium storing computer implementable processor steps for generating within a programmable computer an apparatus in accordance with any of claims 1 to 4.

21. A recording medium in accordance with claim 20, comprising a computer disc.

22. A computer disc in accordance with claim 21, wherein said computer disc comprises an optical, a magneto-optical or magnetic disc.

23. A recording medium in accordance with claim 21, comprising electric signals transferred via the Internet.

24. A method of speech processing utilizing a speech processing apparatus, the method comprising:

storing a plurality of word models in a speech processing apparatus;

storing confidence model data in the speech processing apparatus associating each word model stored in the speech processing apparatus with data indicative of the probability of a value indicative of the goodness of a match being calculated if said word model correctly or incorrectly matches an utterance;

receiving an utterance;

calculating a value indicative of the goodness of the match between a most closely matching word model of said plurality of word models stored in the speech processing apparatus and the received utterance; and

determining and outputting a confidence score indicative of a calculated posterior probability that a received utterance has been correctly matched to a word model given that the match resulted in the calculation of said value indicative of goodness of match utilizing said value and the stored confidence model data associated with said word model,

wherein said calculated value indicative of the goodness of a match comprises data indicative of the ratio of a match score indicative of the correspondence between an utterance and a most closely matching word model and a match score for the correspondence between an utterance and a different word model.

25. A method in accordance with claim 24, wherein a plurality of values indicative of the goodness of a match are calculated, said storing confidence model data comprising storing data associating each word model with data indicative of the probability of each of said values being calculated if said word model correctly or incorrectly matches an utterance and wherein said confidence score is calculated utilizing said plurality of calculated values and said data associated with said word models.

26. A method in accordance with claim 24, wherein said calculated value indicative of the goodness of the match comprises a determined match score indicative of the correspondence between a received utterance and a stored word model.

27. A method in accordance with claim 24, wherein said different word model comprises the second most closely matching word model to an utterance.

28. A method in accordance with claim 24, wherein said different word model comprises the most closely matching word model to an utterance that is representative of a different word to said most closely matching word model.

29. A method in accordance with claim 24, wherein said storing confidence model data comprises storing data indicative of probability density functions of a value being output if an associated word model correctly or incorrectly matches an utterance.

30. A recording medium storing computer implementable processor steps for causing a programmable computer to perform a method in accordance with any of claims 24–

26 and 27–

29.

31. A recording medium in accordance with claim 30, comprising a computer disc.

32. A computer disc in accordance with claim 31, wherein said computer disc comprises an optical, a magneto-optical or magnetic disc.

33. A recording medium in accordance with claim 31, comprising electric signals transferred via the Internet.

34. A speech processing apparatus comprising:

storage means for storing a plurality of word models;

receiving means for receiving an utterance; and

matching means for determining which of a plurality of word models stored in said storage means most closely matches an utterance received by said receiving means, said matching means being arranged to output a value indicative of the goodness of each match between a most closely matching word model and an utterance received by said receiving means;

characterized by further comprising;

association means associating each word model stored in said storage means with data indicative of the probability of a said value indicative of the goodness of match being output if said word model correctly or incorrectly matches an utterance; and

determining means for determining and outputting a confidence score indicative of a calculated posterior probability that a received utterance has been correctly matched to a word model given that the match resulted in the output of a particular value indicative of goodness of match, said determining means being arranged to calculate said confidence score for a match utilizing said value output by said matching means and said data associated with the word model matched to utterances by said association means.