Custom language models for audio content

• US 8,447,608 B1
• Filed: 12/10/2008
• Issued: 05/21/2013
• Est. Priority Date: 12/10/2008
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Systems Incorporated (Adobe Inc.)
• Inventors: Chang, Walter W., Welch, Michael J.
• Primary Examiner(s): Neway, Samuel G
• Application Number:
  US12/332,297
• Time in Patent Office: 1,623 Days
• Field of Search: 704/9, 704231-257
• US Class Current: 704/257
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/242 Dictionaries
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 2015/228 of application context

Abstract

This specification describes technologies relating to generating custom language models for audio content. In some implementations, a computer-implemented method is provided that includes the actions of receiving a collection of source texts; identifying a type from a collection of types for each source text, each source text being associated with a particular type; generating, for each identified type, a type-specific language model using the source texts associated with the respective type; and storing the language models.

Claims

1. A computer-implemented method comprising:

receiving a collection of source texts;

identifying a type from a collection of types for each source text, each source text being associated with a particular type;

generating, by data processing apparatus, a type-specific language model for each identified type using the source texts associated with the respective type;

storing the language models in a computer-readable medium;

receiving an audio source file to be processed, the audio source file having a particular type;

selecting a particular weighted combination of the language models, wherein weighting the particular weighted combination of the language models depends on the type of the audio source file to be processed; and

generating, by data processing apparatus, a text file from the audio source file based on the selected weighted combination of language models.

2. The method of claim 1, further comprising:

processing one or more source texts to extract particular content portions.

3. The method of claim 2, where the one or more source texts are structured documents and the processing includes removing tags indicating document structure.

4. The method of claim 3, where the one or more source texts are in XML format.

5. The method of claim 1, where each source text includes an identifier of the particular type associated with the source text.

6. The method of claim 1, further comprising:

using the weighted combination of language models to generate the text file from the audio source file using speech-to-text processing; and

storing the text file.

7. The method of claim 6, further comprising:

filtering the received audio source file to remove noise prior to generating the text file.

8. The method of claim 6, further comprising:

receiving an input identifying a type of the collection of types associated with the audio source file; and

selecting the weighted combination of language models according to the received input.

9. The method of claim 6, wherein the weighted combination of language models includes a default language model generated using a second collection of text.

10. The method of claim 1, further comprising:

evaluating the generated language models based on a word error rate.

11. The method of claim 10, further comprising:

adjusting the generated language models based on the evaluation of the generated language models; and

storing the adjusted language models.

12. The method of claim 1, where the particular type is based at least in part on a genre of one of a group of cinematic, television, musical, or literary compositions.

13. The method of claim 1, where generating each type-specific language model includes training the type-specific language model using the particular source texts of the type.

14. A computer program product, encoded on a non-transitory computer-readable medium, operable to cause data processing apparatus to perform operations comprising:

receiving a collection of source texts;

identifying a type from a collection of types for each source text, each source text being associated with a particular type;

generating, for each identified type, a type-specific language model using the source texts associated with the respective type;

storing the language models;

receiving an audio source file to be processed, the audio source file having a particular type;

selecting a particular weighted combination of the language models, wherein weighting the particular weighted combination of the language models depends on the type of the audio source file to be processed; and

generating a text file from the audio source file based on the selected weighted combination of language models.

15. The computer program product of claim 14, further operable to perform operations comprising:

processing one or more source texts to extract particular content portions.

16. The computer program product of claim 15, where the one or more source texts are structured documents and the processing includes removing tags indicating document structure.

17. The computer program product of claim 16, where the one or more source texts are in XML format.

18. The computer program product of claim 14, where each source text includes an identifier of the particular type associated with the source text.

19. The computer program product of claim 14, further operable to perform operations comprising:

using the weighted combination of language models to generate the text file from the audio source file using speech-to-text processing; and

storing the text file.

20. The computer program product of claim 19, further operable to perform operations comprising:

filtering the received audio source file to remove noise prior to generating the text file.

21. The computer program product of claim 19, further operable to perform operations comprising:

receiving an input identifying a type of the collection of types associated with the audio source file; and

selecting the weighted combination of language models according to the received input.

22. The computer program product of claim 19, wherein the weighted combination of language models includes a default language model generated using a second collection of text.

23. The computer program product of claim 14, further operable to perform operations comprising:

evaluating the generated language models based on a word error rate.

24. The computer program product of claim 23, further operable to perform operations comprising:

adjusting the generated language models based on the evaluation of the generated language models; and

storing the adjusted language models.

25. The computer program product of claim 14, where the particular type is based at least in part on a genre of one of a group of cinematic, television, musical, or literary compositions.

26. The computer program product of claim 14, where generating each type-specific language model includes training the type-specific language model using the particular source texts of the type.

27. A system comprising:

a processor and a memory operable to perform operations including;

receiving a collection of source texts;

identifying a type from a collection of types for each source text, each source text being associated with a particular type;

generating, for each identified type, a type-specific language model using the source texts associated with the respective type;

storing the language models;

receiving an audio source file to be processed, the audio source file having a particular type;

selecting a particular weighted combination of the language models, wherein weighting the particular weighted combination of the language models depends on the type of the audio source file to be processed; and

generating a text file from the audio source file based on the selected weighted combination of language models.

28. The system of claim 27, further operable to perform operations comprising:

processing one or more source texts to extract particular content portions.

29. The system of claim 28, where the one or more source texts are structured documents and the processing includes removing tags indicating document structure.

30. The system of claim 29, where the one or more source texts are in XML format.

31. The system of claim 27, where each source text includes an identifier of the particular type associated with the source text.

32. The system of claim 27, further operable to perform operations comprising:

using the weighted combination of language models to generate the text file from the audio source file using speech-to-text processing; and

storing the text file.

33. The system of claim 32, further operable to perform operations comprising:

filtering the received audio source file to remove noise prior to generating the text file.

34. The system of claim 32, further operable to perform operations comprising:

receiving an input identifying a type of the collection of types associated with the audio source file; and

selecting the weighted combination of language models according to the received input.

35. The system of claim 32, wherein the weighted combination of language models includes a default language model generated using a second collection of text.

36. The system of claim 27, further operable to perform operations comprising:

evaluating the generated language models based on a word error rate.

37. The system of claim 36, further operable to perform operations comprising:

adjusting the generated language models based on the evaluation of the generated language models; and

storing the adjusted language models.

38. The system of claim 27, where the particular type is based at least in part on a genre of one of a group of cinematic, television, musical, or literary compositions.

39. The system of claim 27, where generating each type-specific language model includes training the type-specific language model using the particular source texts of the type.

40. A method performed by a computer programmed to provide one or more language models, the method comprising:

receiving a collection of source texts;

identifying a type from a collection of types for each source text, each source text being associated with a particular type;

generating, by the computer, a type-specific language model for each identified type using the source texts associated with the respective type;

storing the language models at the computer;

receiving an audio source file to be processed, the audio source file having a particular type;

selecting a weighted combination of the language models, wherein the weights are determined based on the type of the audio source file to be processed; and

generating, by the computer, a text file from the audio source file based on the selected weighted combination of language models.