Speech recognition using final decision based on tentative decisions

• US 5,745,873 A
• Filed: 03/21/1997
• Issued: 04/28/1998
• Est. Priority Date: 05/01/1992
• Status: Expired due to Fees

• Current Assignee: Massachusetts Institute of Technology
• Original Assignee: Massachusetts Institute of Technology
• Inventors: Duchnowski, Paul, Braida, Louis D.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): COLLINS, ALPHONSO
• Application Number:
  US08/821,509
• Time in Patent Office: 403 Days
• Field of Search: 395/2.09, 395/2.31, 395/2.4, 395/2.52, 395/2.56, 395/2.65
• US Class Current: 704/222
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning

Abstract

A method for recognizing speech elements (e.g., phones) in utterances includes the following steps. Based on acoustic frequency, at least two different acoustic representatives are isolated for each of the utterances. From each acoustic representative, tentative decision information on the speech element in the corresponding utterance is derived. A final decision on the speech element in the utterance is then generated, based on the tentative decision information derived from more than one of the acoustic representatives.

Claims

1. A method for making a decision on a speech element in a signal representing an utterance, from among a set of possible speech elements, comprising

generating one representation of said utterance based on one range of acoustic frequencies,

generating at least one separate representation of said utterance based on a different range of acoustic frequencies,

deriving one tentative decision on said speech element from the one representation based on the one range of acoustic frequencies, said one tentative decision being at least one of said speech elements,

deriving at least one separate tentative decision on said speech element from the at least one separate acoustic representation based on the different range of acoustic frequencies, said at least one separate tentative decision being at least one of said speech elements, and

generating a final decision on said speech element from said one tentative decision and said at least one separate tentative decision, said final decision being one of said speech elements.

2. The method of claim 1 further comprising

forming each of said one acoustic representation and said at least one separate acoustic representation based on a band of acoustic frequencies.

3. The method of claim 2 wherein said one acoustic representation and said at least one separate acoustic representation are divided into time-based frames.

4. The method of claim 3 wherein said time-based frames are 20 millisecond long.

5. The method of claim 2 wherein different said bands of said acoustic frequencies are non-overlapping.

6. The method of claim 2 wherein there are at least four said bands.

7. The method of claim 1 wherein said steps of deriving one tentative decision and deriving at least one separate tentative decision includes

forming a parameter vector for each of said one acoustic representation and said at least one separate acoustic representation,

quantizing said parameter vector to form a quantized vector, and

mapping each said quantized vector to one of said tentative decisions.

8. The method of claim 7 wherein said parameter vector has at least one parameter which characterizes said representation in time domain.

9. The method of claim 8 wherein said parameter is created using an autocorrelation function.

10. The method of claim 7 wherein said mapping comprises

applying a sub-recognizer to match each said quantized vector to a model of said speech element.

11. The method of claim 1 or 10 adapted for use in recognizing speech elements in utterances which appear in continuous speech, further comprising

dividing said continuous speech into successive time-based frames, and

applying said method to each of said frames.

12. The method of claim 10 wherein said subrecognizer issues one of said tentative decisions on said speech element per said time-based frame.

13. The method of claim 1 or 12 wherein said speech element comprises a phone.

14. The method of claim 1 wherein said final decision is generated by maximizing a product of a relative speech element occurrence frequency and a relative frequency of observing a phone estimate vector, wherein said relative frequency of observing the phone estimate vector is determined from said tentative decisions.

15. The method of claim 1 wherein said step of generating said final decision includes

forming a global observation index based on said tentative decisions derived from said individual acoustic representations, and

mapping a global observation index to said final decision.

16. The method of claim 15 wherein said mapping comprises

applying a global recognizer to match said global observation index to a model of said speech element.

17. The method of claim 10 or 16 wherein said model comprises a left-to-right hidden Markov model.

18. An apparatus for making a decision on a speech element in a signal representing an utterance, from among a set of possible speech elements, comprising

a filter for isolating one acoustic representation of said utterance based on one range of acoustic frequencies,

at least one additional filter for isolating a separate acoustic representation of said utterance based on a different range of acoustic frequencies,

processing channels for deriving one tentative decision on said speech element from said one representation based on the one range of acoustic frequencies, said one tentative decision being at least one of said speech elements, and for deriving at least one separate tentative decision on said speech element from said separate acoustic representation based on the different range of acoustic frequencies, said at least one separate tentative decision being at least one of said speech elements, and

a global recognizer connected to and receiving from said processing channels said one tentative decision and said at least one separate tentative decision on said speech element to generate a final decision on said speech element, said final decision being one of said speech elements.

19. The apparatus of claim 18 wherein said filters isolate acoustic representations of non-overlapping acoustic frequencies.

20. The apparatus of claim 18 wherein there are at least four filters.

21. The apparatus of claim 18 wherein said processing channel comprises

a parameter extractor for generating a parameter vector for each said acoustic representation,

vector quantizer for quantizing said parameter vector to form a quantized vector, and

a sub-recognizer for mapping each said quantized vector to one of said tentative decisions.

22. The apparatus of claim 21 wherein said acoustic representations are divided into time-based frames.

23. The apparatus of claim 22 wherein said time-based frames of speech are 20 millisecond long.

24. The apparatus of claim 23 wherein said sub-recognizer issues one of said tentative decisions on said speech element per said time-based frame.

25. The apparatus of claim 21 wherein said mapping comprises matching each of said quantized vectors to a model of said speech element.

26. The apparatus of claim 25 wherein said parameter is created using autocorrelation function.

27. The apparatus of claim 21 wherein said parameter vector has at least one parameter which characterizes said acoustic representation in time domain.

28. The apparatus of claim 18 wherein said global recognizer generates said final decision by maximizing a product of a relative speech element occurrence frequency and a relative frequency of observing a phone estimate vector, said relative frequency of observing the phone estimate vector being determined from said tentative decisions.

29. The apparatus of claim 18 wherein said global recognizer is designed to form a global observation index based on said tentative decisions derived from said acoustic representations and to generate a final decision by mapping a global observation index to a model of said speech element.

30. The apparatus of claim 24 or 29 wherein said model comprises a left-to-right hidden Markov model.

31. The apparatus of claim 18, 28 or 29 wherein said speech element comprises a phone.