Language model score look-ahead value imparting device, language model score look-ahead value imparting method, and program storage medium

• US 8,682,668 B2
• Filed: 03/27/2009
• Issued: 03/25/2014
• Est. Priority Date: 05/16/2008
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Okabe, Koji, Isotani, Ryosuke, Yamabana, Kiyoshi, Hanazawa, Ken
• Primary Examiner(s): Desir, Pierre-Louis
• Assistant Examiner(s): SERROU, ABDELALI
• Application Number:
  US12/992,760
• Publication Number:
  US 20110191100A1
• Time in Patent Office: 1,824 Days
• Field of Search: 704/251, 704/231, 704/254, 704/255, 704/235, 704/240
• US Class Current: 704/251
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A speech recognition apparatus that performs frame synchronous beam search by using a language model score look-ahead value prevents the pruning of a correct answer hypothesis while suppressing an increase in the number of hypotheses. A language model score look-ahead value imparting device 108 is provided with a word dictionary 203 that defines a phoneme string of a word, a language model 202 that imparts a score of appearance easiness of a word, and a smoothing language model score look-ahead value calculation means 201. The smoothing language model score look-ahead value calculation means 201 obtains a language model score look-ahead value at each phoneme in the word from the phoneme string of the word defined by the word dictionary 203 and the language model score defined by the language model 202 so that the language model score look-ahead values are prevented from concentrating on the beginning of the word.

Claims

1. A language model score look-ahead value imparting device that comprises:

a word dictionary that defines a phoneme string of a word;

a language model that imparts a score of appearance easiness of a word; and

a smoothing language model score look-ahead value calculation unit, implemented by a processor, that obtains a language model score look-ahead value at each phoneme in the word from the phoneme string of the word defined by the word dictionary and the score defined by the language model, prevents the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word, and obtains an updated language model score look-ahead value at each of the phoneme based on the distribution.

2. The language model score look-ahead value imparting device according to claim 1, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value at each phoneme in the word based on an appearance order of the phoneme in the word.

3. The language model score look-ahead value imparting device according to claim 2, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value at each phoneme in the word based on the number of phonemes from a word head to the phoneme.

4. The language model score look-ahead value imparting device according to claim 3, wherein the smoothing language model score look-ahead value calculation unit obtains a language model score look-ahead value which is within a threshold value of the language model score look-ahead value set based on the number of phonemes from a word head to a phoneme.

5. The language model score look-ahead value imparting device according to claim 2, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value at each phoneme in the word based on the number of phonemes of a word that can be traced from the phoneme.

6. The language model score look-ahead value imparting device according to claim 5, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value so that a language model score look-ahead difference value is equally divided by phonemes that can be traced from the phoneme based on the number of phonemes of a word that can be traced from the phoneme.

7. The language model score look-ahead value imparting device according to claim 2, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value at each phoneme in the word based on the number of phonemes of a phoneme string having no branching and including the phoneme.

8. The language model score look-ahead value imparting device according to claim 7, wherein the smoothing language model score look-ahead value calculation unit obtains the language model score look-ahead value so that a language model score look-ahead difference value is equally divided by phonemes having no branching based on the number of phonemes of a phoneme string having no branching and including the phoneme.

9. A speech recognition apparatus that performs frame synchronous beam search using a language model score look-ahead value, the speech recognition apparatus comprising the language model score look-ahead value imparting device according to claim 1.

10. A computer processor implemented language model score look-ahead value imparting method, the method comprising:

obtaining a language model score look-ahead value at each phoneme in a word from a phoneme string of the word defined by a word dictionary and a score defined by a language model;

preventing, by the computer processor, the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word; and

obtaining an updated language model score look-ahead value at each of the phoneme based on the distribution.

11. The language model score look-ahead value imparting method according to claim 10, wherein the language model score look-ahead value at each phoneme in the word is obtained based on an appearance order of the phoneme in the word.

12. The language model score look-ahead value imparting method according to claim 11, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes from a word head to the phoneme.

13. The language model score look-ahead value imparting method according to claim 12, wherein a language model score look-ahead value which is within a threshold value of the language model score look-ahead value set based on the number of phonemes from a word head to a phoneme is obtained.

14. The language model score look-ahead value imparting method according to claim 11, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes of a word that can be traced from the phoneme.

15. The language model score look-ahead value imparting method according to claim 14, wherein the language model score look-ahead value is obtained so that a language model score look-ahead difference value is equally divided by phonemes that can be traced from the phoneme based on the number of phonemes of a word that can be traced from the phoneme.

16. The language model score look-ahead value imparting method according to claim 11, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes of a phoneme string having no branching and including the phoneme.

17. The language model score look-ahead value imparting method according to claim 16, wherein the language model score look-ahead value is obtained so that a language model score look-ahead difference value is equally divided by phonemes having no branching based on the number of phonemes of a phoneme string having no branching and including the phoneme.

18. A speech recognition method that performs frame synchronous beam search using a language model score look-ahead value obtained by the language model score look-ahead value imparting method according to claim 10.

19. A non-transitory computer readable medium that readably records in a computer a language model score look-ahead value imparting program to cause the computer including a memory unit storing a word dictionary to define a phoneme string of a word and a language model to impart a score of appearance easiness of a word to execute a process of:

obtaining a language model score look-ahead value at each phoneme in the word from the phoneme string of the word defined by the word dictionary and the score defined by the language model;

preventing the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word; and

obtaining an updated language model score look-ahead value at each of the phoneme based on the distribution.

20. The non-transitory computer readable medium according to claim 19, wherein the language model score look-ahead value at each phoneme in the word is obtained based on an appearance order of the phoneme in the word in the process.

21. The non-transitory computer readable medium according to claim 20, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes from a word head to the phoneme in the process.

22. The p non-transitory computer readable medium according to claim 21, wherein a language model score look-ahead value which is within a threshold value of the language model score look-ahead value set based on the number of phonemes from a word head to a phoneme is obtained in the process.

23. The non-transitory computer readable medium according to claim 20, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes of a word that can be traced from the phoneme in the process.

24. The non-transitory computer readable medium according to claim 23, wherein the language model score look-ahead value is obtained so that a language model score look-ahead difference value is equally divided by phonemes that can be traced from the phoneme based on the number of phonemes of a word that can be traced from the phoneme in the process.

25. The non-transitory computer readable medium according to claim 20, wherein the language model score look-ahead value at each phoneme in the word is obtained based on the number of phonemes of a phoneme string having no branching and including the phoneme in the process.

26. The non-transitory computer readable medium according to claim 25, wherein the language model score look-ahead value is obtained so that a language model score look-ahead difference value is equally divided by phonemes having no branching based on the number of phonemes of a phoneme string having no branching and including the phoneme in the process.

27. A non-transitory computer readable medium that records in a computer, in a computer-readable way, a speech recognition program to cause the computer to execute a speech recognition process that performs frame synchronous beam search using a language model score look-ahead value obtained by a language model score look-ahead value imparting program recorded in the program storage medium according to claim 19.

28. A language model score look-ahead value imparting device that comprises:

a word dictionary that defines a symbol string of a word indicating an acoustic model string;

a language model that imparts a score of appearance easiness of a word; and

a smoothing language model score look-ahead value calculation unit that obtains a language model score look-ahead value at each phoneme in the word from the symbol string of the word defined by the word dictionary and the score defined by the language model, prevents the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word, and obtains an updated language model score look-ahead value at each of the phoneme based on the distribution.

29. A computer processor implemented language model score look-ahead value imparting method, the method comprising:

obtaining a language model score look-ahead value at each phoneme in a word from a symbol string of the word indicating an acoustic model string defined by a word dictionary and a score defined by a language model;

preventing, by the computer processor, the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word; and

obtaining an updated language model score look-ahead value at each of the phoneme based on the distribution.

30. A non-transitory computer readable medium that readably records in a computer a language model score look-ahead value imparting program to cause the computer including a memory unit storing a word dictionary to define a symbol string of a word indicating an acoustic model string and a language model to impart a score of appearance easiness of a word to execute a process of:

obtaining a language model score look-ahead value at each phoneme in the word from the symbol string of the word defined by the word dictionary and the score defined by the language model;

preventing the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word; and

obtaining an updated language model score look-ahead value at each of the phoneme based on the distribution.

31. A language model score look-ahead value imparting device that comprises:

a word dictionary that defines a phoneme string of a word;

a language model that imparts a score of appearance easiness of a word; and

a smoothing language model score look-ahead value calculation means that obtains a language model score look-ahead value at each phoneme in the word from the phoneme string of the word defined by the word dictionary and the score defined by the language model, prevents the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word, and obtains an updated language model score look-ahead value at each of the phoneme based on the distribution.

32. A language model score look-ahead value imparting device that comprises:

a word dictionary that defines a symbol string of a word indicating an acoustic model string;

a language model that imparts a score of appearance easiness of a word; and

a smoothing language model score look-ahead value calculation means that obtains a language model score look-ahead value at each phoneme in the word from the symbol string of the word defined by the word dictionary and the score defined by the language model, prevents the language model score look-ahead values from concentrating on the beginning of the word by distributing the language model score look-ahead value corresponding to the phoneme at the beginning of the word to one or more phoneme that are not at the beginning of the word, and obtains an updated language model score look-ahead value at each of the phoneme based on the distribution.