System and method for implementing a refined dictionary for speech recognition

• US 20060031070A1
• Filed: 08/03/2004
• Published: 02/09/2006
• Est. Priority Date: 08/03/2004
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.), Sony Electronics Inc. (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.), Sony Electronics Inc. (Sony Group Corp.)
• Inventors: Abrego, Gustavo, Olorenshaw, Lex S.
• Granted Patent: US 7,529,668 B2
• Time in Patent Office: Days
• US Class Current: 704/243
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A system and method for implementing a refined dictionary for speech recognition includes a database analyzer that initially identifies first vocabulary words that are present in a training database and second vocabulary words that are not present in the training database. A relevance module then performs refinement procedures upon the first vocabulary words to produce refined short word pronunciations and refined long word pronunciations that are added to a refined dictionary. A consensus module compares the second pronunciations with calculated plurality pronunciations to identify final consensus pronunciations that are then included in the refined dictionary.

Claims

1. A system for performing a speech recognition dictionary refinement procedure, comprising:

a database analyzer that identifies first vocabulary words that are present in a training database and identifies second vocabulary words that are not present in said training database;

a relevance module that separates first pronunciations of said first vocabulary words into initial short word pronunciations and initial long word pronunciations, said relevance module performing a cumulative threshold procedure upon said initial short word pronunciations to create final short word pronunciations that are added to a refined dictionary, said relevance module performing a standardized distance threshold procedure upon said initial long word pronunciations to create final long word pronunciations that are added to said refined dictionary; and

a consensus module that computes plurality pronunciations for second pronunciations of said second vocabulary words, said consensus module comparing said plurality pronunciations to said second pronunciations to identify final consensus pronunciations that are added to said refined dictionary.

2. The system of claim 1 further comprising a speech recognition engine that uses said refined dictionary.

3. The system of claim 1 wherein said training database includes recorded speech and corresponding text transcriptions of said recorded speech.

4. The system of claim 1 wherein said refined dictionary is implemented as a series of dictionary entries that each include a text vocabulary word and one or more corresponding vocabulary word pronunciations of one or more phonemes.

5. The system of claim 1 wherein said database analyzer performs a training database analysis procedure by searching said training database for said first vocabulary words and said second vocabulary words.

6. The system of claim 1 wherein said relevance module performs a forced alignment procedure upon said first pronunciations to associate database words from said training database with most closely aligned of said first pronunciations.

7. The system of claim 1 wherein said relevance module determines whether said first pronunciations are observed in said training database.

8. The system of claim 7 wherein said relevance module discards any of said first pronunciations that are not observed in said training database.

9. The system of claim 7 wherein said relevance module performs a word length analysis upon said first pronunciations according to pre-defined word length criteria so as to identify said short word pronunciations and said long word pronunciations.

10. The system of claim 1 wherein said relevance module performs said cumulative threshold procedure to retain as said final short word pronunciations only said initial short word pronunciations that cumulatively account for a pre-determined percentage of corresponding pronunciation observations from said training database.

11. The system of claim 10 wherein said relevance module utilizes a short word pruning exception during said cumulative threshold procedure in which any initial short word pronunciation with only a single observation in said training database is discarded.

12. The system of claim 1 wherein said relevance module performs said standardized distance threshold procedure by retaining only said initial long word pronunciations that have a total number of observations that is greater than a pre-determined standardized threshold value.

13. The system of claim 1 wherein said relevance module calculates a standardized distance according to a formula:

Standardized Distance=(Nmax−

Ncurrent)/standard deviation

where said Nmax is a maximum number of pronunciation observations for said first pronunciations in said training dictionary, said Ncurrent is a current number of said pronunciation observations for a current one of said first pronunciations, and said standard deviation is a standard deviation of said first pronunciations.

14. The system of claim 1 wherein said consensus module performs multiple sequence alignment procedures by aligning corresponding phonemes from said second pronunciations for each of said second vocabulary words.

15. The system of claim 1 wherein said consensus module computes said plurality pronunciations on a phoneme-by-phoneme basis by selecting plurality phonemes that each represent a respective most frequent phoneme in a given phoneme position from aligned sets of said second pronunciations.

16. The system of claim 1 wherein said consensus module determines whether said plurality pronunciations are validly represented in said second pronunciations.

17. The system of claim 16 wherein said consensus module adds said plurality pronunciations to said refined dictionary when said plurality pronunciations are validly represented in said second pronunciations.

18. The system of claim 16 wherein said consensus module performs a pair-wise alignment procedure to compare said plurality pronunciations with said second pronunciations if said plurality pronunciations are not validly represented in said second pronunciations.

19. The system of claim 18 wherein said consensus module selects said final consensus pronunciations by identifying said second pronunciations that are most similar to respective ones of said plurality pronunciations.

20. The system of claim 1 wherein said consensus module discards more of said second pronunciations than said relevance module discards of said first pronunciations, said relevance module discarding more of said initial long word pronunciations than of said initial short word pronunciations.

21. A method for performing a speech-recognition dictionary refinement procedure, comprising:

utilizing a database analyzer to identify first vocabulary words that are present in a training database, and to identify second vocabulary words that are not present in said training database;

separating first pronunciations of said first vocabulary words into initial short word pronunciations and initial long word pronunciations by utilizing a relevance module that then performs a cumulative threshold procedure upon said initial short word pronunciations to create final short word pronunciations that are added to a refined dictionary, said relevance module also performing a standardized distance threshold procedure upon said initial long word pronunciations to create final long word pronunciations that are added to said refined dictionary; and

computing plurality pronunciations for second pronunciations of said second vocabulary words by utilizing a consensus module that then compares said plurality pronunciations to said second pronunciations to identify final consensus pronunciations that are added to said refined dictionary.

22. The method of claim 21 further comprising a speech recognition engine that uses said refined dictionary.

23. The method of claim 21 wherein said training database includes recorded speech and corresponding text transcriptions of said recorded speech.

24. The method of claim 21 wherein said refined dictionary is implemented as a series of dictionary entries that each include a text vocabulary word and one or more corresponding vocabulary word pronunciation of one or more phonemes.

25. The method of claim 21 wherein said database analyzer performs a training database analysis procedure by searching said training database for said first vocabulary words and said second vocabulary words.

26. The method of claim 21 wherein said relevance module performs a forced alignment procedure upon said first pronunciations to associate database words from said training database with most closely aligned of said first pronunciations.

27. The method of claim 21 wherein said relevance module determines whether said first pronunciations are observed in said training database.

28. The method of claim 27 wherein said relevance module discards any of said first pronunciations that are not observed in said training database.

29. The method of claim 27 wherein said relevance module performs a word length analysis upon said first pronunciations according to pre-defined word length criteria so as to identify said short word pronunciations and said long word pronunciations.

30. The method of claim 21 wherein said relevance module performs said cumulative threshold procedure to retain as said final short word pronunciations only said initial short word pronunciations that cumulatively account for a pre-determined percentage of corresponding pronunciation observations from said training database.

31. The method of claim 30 wherein said relevance module utilizes a short word pruning exception during said cumulative threshold procedure in which any initial short word pronunciation with only a single observation in said training database is discarded.

32. The method of claim 21 wherein said relevance module performs said standardized distance threshold procedure by retaining only said initial long word pronunciations that have a total number of observations that is greater than a pre-determined standardized threshold value.

33. The method of claim 21 wherein said relevance module calculates a standardized distance according to a formula:

Standardized Distance=(Nmax−

Ncurrent)/standard deviation

where said Nmax is a maximum number of pronunciation observations for said first pronunciations in said training dictionary, said Ncurrent is a current number of said pronunciation observations for a current one of said first pronunciations, and said standard deviation is a standard deviation of said first pronunciations.

34. The method of claim 21 wherein said consensus module performs multiple sequence alignment procedures by aligning corresponding phonemes from said second pronunciations for each of said second vocabulary words.

35. The method of claim 21 wherein said consensus module computes said plurality pronunciations on a phoneme-by-phoneme basis by selecting plurality phonemes that each represent a respective most frequent phoneme in a given phoneme position from aligned sets of said second pronunciations.

36. The method of claim 21 wherein said consensus module determines whether said plurality pronunciations are validly represented in said second pronunciations.

37. The method of claim 36 wherein said consensus module adds said plurality pronunciations to said refined dictionary when said plurality pronunciations are validly represented in said second pronunciations.

38. The method of claim 36 wherein said consensus module performs a pair-wise alignment procedure to compare said plurality pronunciations with said second pronunciations if said plurality pronunciations are not validly represented in said second pronunciations.

39. The method of claim 38 wherein said consensus module selects said final consensus pronunciations by identifying said second pronunciations that are most similar to respective ones of said plurality pronunciations.

40. The method of claim 21 wherein said consensus module discards more of said second pronunciations than said relevance module discards of said first pronunciations, said relevance module discarding more of said initial long word pronunciations than of said initial short word pronunciations.

41. A system for performing a speech-recognition dictionary refinement procedure, comprising:

means for identifying first vocabulary words that are present in a training database, and second vocabulary words that are not present in said training database;

means for separating first pronunciations of said first vocabulary words into initial short word pronunciations and initial long word pronunciations, said means for separating then performing a cumulative threshold procedure upon said initial short word pronunciations to create final short word pronunciations that are added to a refined dictionary, said means for separating also performing a standardized distance threshold procedure upon said initial long word pronunciations to create final long word pronunciations that are added to said refined dictionary; and

means for computing plurality pronunciations for second pronunciations of said second vocabulary words, said means for computing then comparing said plurality pronunciations to said second pronunciations to identify final consensus pronunciations that are added to said refined dictionary.

42. A system for performing a speech-recognition dictionary refinement procedure, comprising:

a database analyzer that identifies first vocabulary words that are present in a training database, and second vocabulary words that are not present in said training database;

a relevance module that generates refined short word pronunciations of said first vocabulary words for a refined dictionary, said relevance module also generating refined long word pronunciations of said first vocabulary words for said refined dictionary; and

a consensus module that generates consensus pronunciations for said refined dictionary based upon plurality pronunciations of said second vocabulary words.