Method for speaker adaptation of speech models recognition scheme using the method and recording medium having the speech recognition method recorded thereon

• US 5,835,890 A
• Filed: 04/09/1997
• Issued: 11/10/1998
• Est. Priority Date: 08/02/1996
• Status: Expired due to Term

• Current Assignee: Nippon Telegraph and Telephone Corporation
• Original Assignee: Nippon Telegraph and Telephone Corporation
• Inventors: Matsui, Tomoko, Furui, Sadaoki
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/835,421
• Time in Patent Office: 580 Days
• Field of Search: 704/231, 704/256, 704/200, 704/232, 704/236, 704/240, 704/245, 704/241, 704/243, 704/244, 704/255, 704/257, 704/242, 702/239
• US Class Current: 704/255
• CPC Class Codes: G10L 15/065 Adaptation

Abstract

In a speaker adaptation method for speech models, input speech is transformed to a feature parameter sequence like a cepstral sequence, and N model sequences of maximum likelihood for the feature parameter sequence are extracted from speaker-independent speech HMMs by an N-best hypothesis extraction method. The extracted model sequences are each provisionally adapted to maximize its likelihood for the feature parameter sequence of the input speech while changing the HMM parameters of each sequence, and that one of the provisionally adapted model sequences which has the maximum likelihood for the feature parameter sequence of the input speech is selected and speech models of the selected sequence are provided as adapted HMMs of the speaker to be recognized.

Claims

1. A speaker adaptation method for speech models whereby speaker-independent speech models, which form a reference model dictionary modeled by model parameters from speech features corresponding to a phoneme, word or similar recognition category in speech of many speakers, are adapted to speech of a speaker to be recognized, said method comprising:

(a) a feature parameter extracting step of extracting a speech feature parameter sequence of input speech of said speaker to be recognized;

(b) a model sequence extracting step of extracting from said reference model dictionary a plurality of hypothetical model sequences estimated to correspond to said feature parameter sequence of said input speech;

(c) a provisional adaptation step of performing provisional adaptation of each of said plurality of extracted hypothetical model sequences by controlling its model parameters in a manner to maximize its likelihood for said feature parameter sequence of said input speech; and

(d) an adapted model selecting step of selecting, as an adapted speech model sequence, at least one of said provisionally adapted hypothetical model sequences on the basis of their likelihoods for said feature parameter sequence of said input speech.

2. The method of claim 1, wherein said adapted model selecting step (d) is a step of selecting that one of said hypothetical model sequences provisionally adapted in said provisional adaptation step (c) which has the maximum likelihood for said feature parameter sequence of said input speech.

3. The method of claim 1 or 2, which repeats at least once:

a step of adapting said speaker-independent speech models on the basis of model parameters of said adapted speech model sequence selected in said adapted model selecting step (d);

a step of performing said provisional adaptation step while finely performing said provisional adaptation of said each extracted hypothetical model sequence with the ties of its transformation parameters loosened more than in the immediately previous provisional adaptation through the use of said adapted speaker-independent speech models; and

a step of re-extracting said adapted model sequence in said adapted model selecting step (d).

4. A speech recognition method which follows said method of claim 3 and includes:

a reference model dictionary adaptation step of adapting all speech models of said reference model dictionary to said model sequence selected in said adapted model selecting step; and

a re-recognition step of extracting from said adapted reference model dictionary a speech model sequence of the maximum likelihood for said feature parameter sequence of said input speech and outputting said extracted speech model as the result of recognition.

5. The speaker adaptation method of claim 4, wherein said re-recognizing step is a step of calculating, by a Viterbi algorithm, a model sequence which has the maximum likelihood function for said input speech data transformed to a time sequence of feature parameters.

6. The method of claim 1 or 2, further comprising:

a reference model dictionary adaptation step of adapting speech models of said reference model dictionary to said provisionally adapted model sequence in said adapted model selecting step (d);

a step of re-selecting said plurality of hypothetical model sequences of the maximum likelihoods for said input speech through the use of said adapted reference model dictionary in said model sequence extracting step (b) and of performing said provisional adaptation step (c) while finely performing said provisional adaptation of said each extracted hypothetical model sequence with the ties of its transformation parameters loosened more than in the immediately previous provisional adaptation; and

a step of selecting, as a re-adapted model sequence, that one of said provisionally adapted hypothetical mode sequences which has the maximum likelihood.

7. A speech recognition method which follows said method of claim 6 and includes a step of re-recognizing said feature parameter sequence of said input speech by the use of speaker-independent speech models of said reference model dictionary adapted finally, extracting one model sequence of the maximum likelihood for said feature parameter sequence, and outputting said extracted model sequence as the result of recognition of said input speech of a speaker to be recognized.

8. A recording medium having recorded thereon the speech recognition method of claim 7.

9. The speaker adaptation method of claim 7, wherein said re-recognizing step is a step of calculating, by a Viterbi algorithm, a model sequence which has the maximum likelihood function for said input speech data transformed to a time sequence of feature parameters.

10. The method of claim 2, further comprising a step of repeating said provisional adaptation step in succession a predetermined number of times for fine provisional adaptation with the ties of said transformation parameters loosened more than in the immediately previous provisional adaptation.

11. The method of claim 2, further comprising a step of repeating said provisional adaptation step in succession a predetermined number of times for fine provisional adaptation with the ties of said transformation parameters loosened more than in the immediately previous provisional adaptation.

12. The method of claim 1 or 2, wherein said model sequence extracting step (b) is a step of obtaining a predetermined number of model sequences which has the maximum likelihood for said feature parameter sequence of said input speech.

13. The method of claim 1 or 2, wherein said adapted model selecting step (d) is a step of selecting said speech models as the sum of adapted speech models corresponding to said extracted model sequences after being provisionally adapted in said provisional adaptation step (c), said adapted speech models being weighted corresponding to their likelihoods for said feature parameter sequence of said input speech.

14. The method of claim 1 or 2, wherein said speech models are speech HMMs defined by statistical parameters.

15. The method of claim 14, wherein said provisional adaptation step (c) is a step of adapting said hypothetical models in a manner to maximize their likelihoods for said feature parameter sequence of said input speech by controlling the mean values of Gaussian distributions which are at least one of said statistical parameters.

16. A speech recognition method which follows said method of claim 1 or 2 and includes a step of outputting, as a result of recognition of said input speech of said speaker to be recognized, a recognition category corresponding to said model sequence selected in said adapted model selecting step (d).

17. A recording medium having recorded thereon the speech recognition method of claim 16.

18. A speech recognition method which follows said method of claim 2 and includes:

a reference model dictionary adaptation step of adapting all speech models of said reference model dictionary to said model sequence selected in said adapted model selecting step; and

a re-recognition step of extracting from said adapted reference model dictionary a speech model sequence of the maximum likelihood for said feature parameter sequence of said input speech and outputting said extracted speech model as the result of recognition.

19. The speaker adaptation method of claim 18, wherein said re-recognizing step is a step of calculating, by a Viterbi algorithm, a model sequence which has the maximum likelihood function for said input speech data transformed to a time sequence of feature parameters.

20. A recording medium having recorded thereon the speech recognition method of claim 19.