Template generation method in a speech recognition system

• US 4,751,737 A
• Filed: 11/06/1985
• Issued: 06/14/1988
• Est. Priority Date: 11/06/1985
• Status: Expired due to Term

• Current Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Lindsley, Brett L., Gerson, Ira A.
• Primary Examiner(s): Shoop, Jr., William M.
• Assistant Examiner(s): Ip, Paul
• Application Number:
  US06/795,562
• Time in Patent Office: 951 Days
• Field of Search: 381/41, 381/42, 381/43, 381/44, 381/45, 381/46, 381/50, 381/51, 381/36, 381/39, 364/513.5, 364/521, 364/410, 364/487, 364/513, 340/825.23, 379/88, 379/89, 379/199, 379/354, 379/355
• US Class Current: 704/243
• CPC Class Codes: G10L 15/063 Training

Abstract

Disclosed is a method for generating word templates for a speech recognition system. It is used where speech is represented by data in frames of equal time intervals. The method includes generating an interim template, generating a time alignment path between the interim template and a token, mapping frames from the interim template and the token along the time alignment path onto an averaged time axis, and combining data associated with the mapped frames to produce composite frames representative of the final word template. The method realizes advantages of reduced memory usage and a realistic data average from each contributing averaged word.

Claims

1. In a speech recognition system, wherein speech is represented by data in frames of equal time intervals, a method for generating a final word template from a plurality of tokens, comprising the steps of:

(a) forming an interim template representative of at least one token;

(b) generating a time alignment path between said interim template and an additional token;

(c) mapping frames from said interim template and said additional token along said time alignment path onto an averaged time axis; and

(d) combining data associated with said mapped frames to produce composite frames representative of the final word template.

2. A method for generating a final word template, according to claim 1, wherein forming said interim template includes the steps of:

generating a time alignment path between a first and second token;

mapping frames from said first and second token along said time alignment path onto an averaged time axis; and

combining data associated with said mapped frames to produce composite frames representative of the interim template.

3. A method for generating a final word template, according to claim 1, wherein combining data comprises averaging the data such that the resultant data is approximately equally contributed to by each token the final template represents.

4. A method for generating a final word template, according to claim 1, including the steps of comparing the data between the interim template and the additional token to produce a distance measure, and comparing said distance measure to a predetermined distance measure.

5. A method for generating a final word template, according to claim 1, including the step of weighting the data representing said interim template.

6. A method for generating a final word template, according to claim 1, including the step of repetitively averaging data representing tokens with data representing said interim template.

7. A method for generating a final word template, according to claim 1, including the step of accumulating values assigned to the direction taken along said time alignment path to produce said composite frames.

8. A method for generating a final word template, according to claim 1, wherein said combining includes the step of accumulating the data representing said interim template and the data representing said additional token.

9. A method for generating a final word template, according to claim 1, including the step of normalizing a sum of accumulated data to produce said composite frames.

10. In a speech recognition system, wherein speech is represented by data in frames of equal time intervals, a method for generating a final word template from a plurality of tokens, comprising the steps of:

(a) forming an interim template representative of at least one token;

(b) generating a time alignment path between said interim template and an additional token;

(c) weighting the data representing said interim template proportional to the number of tokens said interim template represents; and

(d) combining frames from said interim template with frames from said additional token, dependent upon the number of tokens which the interim template represents, to produce output frames representative of the final word template.

11. A method for generating a final word template, according to claim 10, including the step of mapping frames from said interim template and said token along said time alignment path onto an averaged time axis.

12. A method for generating a final word template, according to claim 10, wherein forming said interim template includes the steps of

generating a time alignment path between a first and second token;

mapping frames from said first and second token along said time alignment path onto an averaged time axis; and

combining data associated with said mapped frames to produce composite frames representative of the interim template.

13. A method for generating a final word template, according to claim 10, including the steps of comparing the data between the interim template and the additional token to produce a distance measure and comparing said distance measure to a predetermined distance measure.

14. A method for generating a final word template, according to claim 10, wherein said interim template includes an approximately equal contribution of data from each token said interim template represents.

15. A method for generating a final word template according to claim 10, wherein weighting the data representing said interim template includes the step of assigning at least one value to each direction taken along said time alignment path.

16. A method for generating a final word template, according to claim 10, wherein said combining includes the step of accumulating the weighted data representing said interim template and the data representing said additional token.

17. A method for generating a final word template, according to claim 10, including the step of normalizing a sum of accumulated data representing averaged frames.

18. In a speech recognition system, wherein speech is represented by data in frames of equal time intervals, an arrangement for generating a final word template from a plurality of tokens, including:

(a) means for forming an interim template representative of at least one token;

(b) means for generating a time alignment path between said interim template and an additional token;

(c) means for mapping frames from said interim template and said additional token along said time alignment path onto an averaged time axis; and

(d) means for combining data associated with said mapped frames to produce composite frames representative of a final word template.

19. An arrangement, according to claim 18, further comprising means for normalizing an accumulation of combined data to form said composite frames.

20. An arrangement, according to claim 18, including means for comparing the data between the interim template and the additional token to produce a distance measure, and means for comparing said distance measure to a predetermined distance measure.

21. An arrangement, according to claim 18, wherein said interim template includes an approximately equal contribution of data from each token said interim template represents.

22. An arrangement, according to claim 18, including means for weighting the data representing said interim template.

23. An arrangement, according to claim 18, including means for repetitively averaging data representing tokens with data representing said interim template.

24. In a speech recognition system, wherein speech is represented by data in frames of equal time intervals, an arrangement for generating a final word template from a plurality of tokens, including:

(a) means for forming an interim template representative of at least one token;

(b) means for generating a time alignment path between said interim template and an additional token;

(c) means for weighting the data representing said interim template proportional to the number of tokens said interim template represents; and

(d) means for combining frames from said interim template with frames from said additional token, dependent upon the number of tokens which the interim template represents, to produce output frames representative of the final word template.

25. An arrangement, according to claim 24, further comprising means for normalizing an accumulation of combined data to form said composite frames.

26. An arrangement, according to claim 24, including means for comparing the data between the interim template and the additional token to produce a distance measure and means for comparing said distance measure to a predetermined distance measure.

27. An arrangement, according to claim 24, wherein said interim template includes an approximately equal contribution of data from each token said interim template represents.

28. An arrangement, according to claim 24, including means for mapping frames from said interim template and said additional token along said time alignment path onto an averaged time axis.

29. An arrangement, according to claim 24, including means for repetitively averaging data representing tokens with data representing said interim template.