Improving speech recognition through text-based linguistic post-processing

• US 6,064,957 A
• Filed: 08/15/1997
• Issued: 05/16/2000
• Est. Priority Date: 08/15/1997
• Status: Expired due to Term

• Current Assignee: General Electric Company
• Original Assignee: General Electric Company
• Inventors: Brandow, Ronald Lloyd, Strzalkowski, Tomasz
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US08/911,247
• Time in Patent Office: 1,005 Days
• Field of Search: 704/231, 704/243, 704/244, 704/255, 704/256
• US Class Current: 704/235
• CPC Class Codes: G06F 40/194 Calculation of difference between files
• CPC Class Codes: G06F 40/232 Orthographic correction, e.g. spell checking or vowelisation
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

The present invention discloses a method and system for improving speech recognition. In this invention, there is a training phase where text data generated from a speech recognition system is collected and aligned with a corresponding true transcription of the speech recognition text data. A preliminary set of correction rules are generated and observed against a corpus of fully verified text data. Rules that are applicable are validated, while invalid rules are updated. The updated rules are then applied to the parallel sample of speech recognition text data and corresponding text data, as well as the corpus of text data. The rules are examined again to determine their validity. This process continues until all of the rules have been validated or until no further progress is made. The finalized correction rules are then put into a production phase.

Claims

1. A method for improving speech recognition, comprising the steps of:

collecting text data generated from a speech recognition system;

collecting a corresponding true transcription of the speech recognition text data;

aligning the text data generated from the speech recognition system with the corresponding true transcription of text data, wherein the aligning is text-based;

generating a plurality of correction rules from differences in alignment between the text data generated from the speech recognition system and the corresponding true transcription of text data; and

applying the plurality of correction rules to new text data generated from a speech recognition system.

2. The method according to claim 1, wherein the step of aligning comprises aligning the text data generated from the speech recognition system with the corresponding true transcription of text data on a word level.

3. The method according to claim 2, wherein the step of aligning comprises examining any differences in alignment between the speech recognition text data and the corresponding true transcription of text data.

4. The method according to claim 1, wherein the plurality of correction rules comprise a plurality of context-free rules and a plurality of context-sensitive rules.

5. The method according to claim 4, wherein the plurality of correction rules further comprise a plurality rules containing non-terminal symbols.

6. The method according to claim 1, further comprising the step of validating each of the plurality of correction rules.

7. The method according to claim 6, wherein the step of validating each of the plurality of correction rules comprises the steps of:

specifying a string within the text data generated from the speech recognition system;

applying a correction rule to the specified string; and

determining the number of occurrences that the applied correction rule is supported in the corresponding true transcription of text data.

8. The method of claim 7, further comprising the steps of:

collecting a corpus of fully verified text data;

applying the correction rule to the corpus of fully verified text data; and

determining the applicability of the correction rule across the corpus of fully verified text data.

9. The method according to claim 8, further comprising the steps of:

revising the correction rule if not supported by the corpus of fully verified text data; and

revalidating the revised correction rule.

10. A system for improving speech recognition, comprising:

a text aligner for aligning text data generated from a speech recognition system with a corresponding true transcription of the speech recognition text data, wherein the aligning is text-based;

a rule generator coupled to the text aligner for generating a plurality of correction rules from differences in alignment between the speech recognition text data and the corresponding true transcription of text data; and

a rule administrator for applying the plurality of correction rules to new text data generated from a speech recognition system.

11. The system according to claim 10, wherein the text aligner aligns the speech recognition text data with the corresponding true transcription of text data on a word level.

12. The system according to claim 11, wherein the text aligner examines any differences in alignment between the speech recognition text data and the corresponding true transcription of text data.

13. The system according to claim 10, wherein the plurality of correction rules comprise a plurality of context-free rules and a plurality of context-sensitive rules.

14. The system according to claim 13, wherein the plurality of correction rules further comprise a plurality of rules containing non-terminal symbols.

15. The system according to claim 10, further comprising a rule validator for validating each of the plurality of correction rules.

16. The system according to claim 15, wherein the rule validator comprises:

means for specifying a string within the speech recognition text data;

means for applying a correction rule to the specified string; and

means for determining the number of occurrences that the applied rule is supported in the corresponding true transcription of text data.

17. The system of claim 16, further comprising:

a corpus of fully verified text data;

means for applying the correction rule to the corpus of fully verified text data; and

means for determining the applicability of the correction rule across the corpus of fully verified text data.

18. The system according to claim 17, further comprising:

means for revising the correction rule if not supported by the corpus of fully verified text data; and

means for revalidating the revised correction rule.