Language model for use in speech recognition

• US 7,584,102 B2
• Filed: 11/15/2002
• Issued: 09/01/2009
• Est. Priority Date: 11/15/2002
• Status: Expired due to Fees

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: ScanSoft, Inc. n/k/a Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Hwang, Kwangil, Fieleke, Eric
• Primary Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US10/294,901
• Publication Number:
  US 20040098263A1
• Time in Patent Office: 2,482 Days
• Field of Search: 704/235, 704/255, 704/242, 704/256
• US Class Current: 704/256
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

Building a language model for use in speech recognition includes identifying without user interaction a source of text related to a user. Text is retrieved from the identified source of text and a language model related to the user is built from the retrieved text.

Claims

1. A method for building a user-specific language model for use in speech recognition by a user in which at least one computer system accesses instructions from computer storage and executes the instructions comprising steps of:

identifying without user interaction one or more sources of non-verbal, electronically embedded, and previously created text that is related to the user and that is located only on a user'"'"'s personal storage, the identified sources of text comprising at least one email sent by the user;

retrieving text from the identified one or more sources of text; and

building a user-specific language model related to the user using only the retrieved text from the identified one or more sources of previously created text that is related to the user and is located only on the user'"'"'s personal storage.

2. The method as in claim 1 further comprising:

scanning the emails to identify words from the emails; and

determining which of the scanned words to use to build the language model.

3. The method as in claim 2, wherein scanning the emails includes:

identifying at least one word in the email sent by the user to add to a vocabulary of the language model.

4. The method as in claim 3, wherein identifying the word in the email sent by the user to add to the vocabulary of the language model includes:

applying at least one filter rule to disregard text from at least one of;

header information of the email and address information of the email; and

upon applying the filter rule, selecting at least one word in remaining text of the email.

5. The method as in claim 3, wherein identifying the word in the email sent by the user to add to the vocabulary of the language model includes:

applying at least one filter rule to disregard reply text in the email message, the reply text copied into the email message from at least one previously transmitted email message; and

upon applying the filter rule, selecting at least one word in remaining text of the email.

6. The method as in claim 1 wherein identifying the one or more sources of text includes identifying without user interaction documents authored by the user.

7. The method as in claim 1 wherein identifying the one or more sources of text includes identifying without user interaction documents stored in a particular location on a computer.

8. The method as in claim 1 wherein identifying the one or more sources of text includes identifying without user interaction contacts information of the user.

9. The method as in claim 8 wherein retrieving the text includes retrieving names from the user'"'"'s contacts information and further comprising:

scanning the names to identify names that are not included in a vocabulary of words; and

adding without user interaction the identified names to the vocabulary.

10. The method as in claim 1 further comprising:

scanning the retrieved text to identify words from the text that are not included in a vocabulary of words; and

adding without user interaction the identified words to the vocabulary.

11. The method as in claim 1 wherein building the language model includes calculating at least one of unigrams, bigrams, and trigrams using the retrieved text.

12. The method as in claim 1 further comprising performing the identifying, retrieving and building as a part of a user'"'"'s enrollment process.

13. The method as in claim 1, wherein retrieving text from the identified one or more sources of text includes:

retrieving at least one document containing at least a required number of characters.

14. A method for building a user-specific language model for use in speech recognition by a user in which at least one computer system accesses instructions from computer storage and executes the instructions comprising steps of:

identifying without user interaction a source of text created by the user, the identified source of text comprising at least one email sent by the user;

retrieving text from the identified source of text;

building a user-specific language model related to the user using only the retrieved text;

scanning the retrieved text to identify words from the text that are not included in a vocabulary of words;

displaying the identified words to the user;

prompting the user to select words to be included in the vocabulary; and

adding words selected by the user to the vocabulary.

15. A computer program for building a user-specific language model for use in speech recognition by a user, the computer program described according to computer code stored on a computer readable storage medium, the computer code comprising:

an identifying code segment that causes a computer to identify without user interaction one or more sources of non-verbal, electronically embedded, and previously created text that is related to the user and that is located only on a user'"'"'s personal storage;

a retrieving code segment that causes the computer to retrieve text from the identified one or more sources of text, the identified sources of text comprising at least one email sent by the user; and

a building code segment that causes the computer to build a user-specific language model related to the user using only the retrieved text from the identified one or more sources of previously created text that is related to the user and that is located only on the user'"'"'s personal storage.

16. The computer program of claim 15 further comprising:

a scanning code segment that causes the computer to scan the emails to identify words from the emails; and

a determining code segment that causes the computer to determine which of the scanned words to use to build the language model.

17. The computer program of claim 15 wherein the identifying code segment causes the computer to identify without user interaction documents authored by the user.

18. The computer program of claim 15 wherein the identifying code segment causes the computer to identify without user interaction documents stored in a particular location on a computer.

19. The computer program of claim 15 wherein the identifying code segment causes the computer to identify without user interaction contacts information of the user.

20. The computer program of claim 19 wherein the retrieving code segment causes the computer to retrieve names from the user'"'"'s contacts information and further comprising:

a scanning code segment that causes the computer to scan the names to identify names that are not included in a vocabulary of words; and

an adding code segment that causes the computer to add without user interaction the identified names to the vocabulary.

21. The computer program of claim 15 further comprising:

a scanning code segment that causes the computer to scan the retrieved text to identify words from the text that are not included in a vocabulary of words; and

an adding code segment that causes the computer to add without user interaction the identified words to the vocabulary.

22. The computer program of claim 15 wherein the building code segment causes the computer to calculate at least one of unigrams, bigrams, and trigrams using the retrieved text.

23. The computer program of claim 15 further comprising a performing code segment that causes the computer to perform the identifying code segment, the retrieving code segment and the building code segment as a part of a user'"'"'s enrollment process.

24. A computer program for building a user-specific language model for use in speech recognition by a user, the computer program described according to computer code stored on a computer readable storage medium, the computer code comprising:

an identifying code segment that causes a computer to identify without user interaction a source of text created by the user, the source of text comprising at least one email sent by the user;

a retrieving code segment that causes the computer to retrieve text from the identified source of text;

a building code segment that causes the computer to build a user-specific language model related to the user using only the retrieved text;

a scanning code segment that causes the computer to scan the retrieved text to identify words from the text that are not included in a vocabulary of words;

a displaying code segment that causes the computer to display the identified words to the user;

a prompting code segment that causes the computer to prompt the user to select words to be included in the vocabulary; and

an adding code segment that causes the computer to add words selected by the user to the vocabulary.

25. A system for building a user-specific language model for use in speech recognition by a user, comprising:

means for identifying without user interaction one or more sources of non-verbal, electronically embedded, and previously created text that is related to the user and that is located only on a user'"'"'s personal storage, the identified sources of text comprising at least one email sent by the user;

means for retrieving text from the identified one or more sources of text; and

means for building a user-specific language model related to the user using only the retrieved text from the identified one or more sources of previously created text that is related to the user and that is located only on the user'"'"'s personal storage.

26. A method for building a user-specific language model for use in speech recognition by a user in which at least one computer system accesses instructions from computer storage and executes the instructions comprising steps of:

identifying without user interaction sources of non-verbal and electronically embedded text related to the user and located only on a user'"'"'s personal storages, the identified sources of text comprising at least one email sent by the user;

retrieving text from the identified sources of text; and

building a language model related to the user using only the retrieved text from the identified sources of text related to that user and located only on the user'"'"'s personal storage.

27. The method as in claim 26 wherein identifying the sources of text includes identifying without user interaction documents authored by the user.

28. The method as in claim 26 wherein identifying the sources of text includes identifying without user interaction documents stored in a particular location on a computer.

29. The method as in claim 26 further comprising:

scanning the retrieved text to identify words from the text that are not included in a vocabulary of words; and

adding without user interaction the identified words to the vocabulary.

30. A method for building a user-specific language model for use in speech recognition by a user in which at least one computer system accesses instructions from computer storage and executes the instructions comprising steps of:

identifying without user interaction sources of text related to the user, the identified sources of text comprising at least one email sent by the user;

retrieving text from the identified sources of text;

building a language model related to the user using only the retrieved text;

scanning the retrieved text to identify words from the text that are not included in a vocabulary of words;

displaying the identified words to the user;

prompting the user to select words to be included in the vocabulary; and

adding words selected by the user to the vocabulary.