Apparatus and method for performing model estimation utilizing a discriminant measure

• US 5,970,239 A
• Filed: 08/11/1997
• Issued: 10/19/1999
• Est. Priority Date: 08/11/1997
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Padmanabhan, Mukund, Bahl, Lalit Rai
• Primary Examiner(s): Teska, Kevin J.
• Assistant Examiner(s): Broda, Samuel
• Application Number:
  US08/908,120
• Time in Patent Office: 799 Days
• Field of Search: 364/578, 704/231, 704/236
• US Class Current: 704/245
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

Method for performing acoustic model estimation to optimize classification accuracy on speaker derived feature vectors with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond comprises: (a) initializing an acoustic model for each phone; (b) evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the model for the phone assigns to feature vectors from the phone and a second component is defined as a probability that the model for the phone assigns to feature vectors from other phones; (c) adapting the model for selected phones so as to increase the first component for the phone or decrease the second component for the phone, the adapting step yielding a new model for each selected phone; (d) evaluating the merit of the new models for each phone adapted in step (c) utilizing the two component measure; (e) comparing results of the evaluation of step (b) with results of the evaluation of step (d) for each phone, and if the first component has increased or the second component has decreased, the new model is kept for that phone, else the model originally initialized is kept; (f) estimating parameters associated with each model kept for each phone in order to optimize the function; and (g) evaluating termination criterion to determine if the parameters of the models are optimized.

Claims

1. Apparatus for performing acoustic model estimation in order to optimize classification accuracy on feature vectors derived from a speaker with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond, the apparatus comprising:

means for initializing an acoustic model for each class;

first means for evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from the phone and a second component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from other phones;

means for adapting the acoustic model for selected phones so as to one of increase the first component of the discriminant measure for the phone and decrease the second component of the discriminant measure for the phone, the adapting means yielding a new acoustic model for each selected phone;

second means for evaluating the merit of the new acoustic models for each phone adapted by the adapting means utilizing the two component discriminant measure;

means for comparing results obtained by the first evaluating means with results obtained by the second evaluating means for each phone, and if one of the first component of the discriminant measure has increased and the second component of the discriminant measure has decreased, then the new acoustic model is kept for that phone, else the acoustic model originally initialized is kept;

means for estimating parameters associated with each acoustic model kept for each phone in order to substantially optimize the objective function; and

third means for evaluating termination criterion to determine if the parameters of the acoustic models are substantially optimized.

2. The apparatus of claim 1, further comprising means for sequentially repeating the functions respectively performed by the first evaluating means, the adapting means, the second evaluating means, the comparing means, the estimating means and the third evaluating means if the termination criterion has not been substantially satisfied.

3. The apparatus of claim 1, wherein the first component of the two component discriminant measure is represented as:

##EQU10## where x_t represents the feature vectors and T₁ represents a normalizing factor and P_c^l (x_t) is represented as;

##EQU11## where M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

4. The apparatus of claim 1, wherein the second component of the two component discriminant measure is represented as:

##EQU12## where x_t represents the feature vectors and T₂ represents a normalizing factor and where P_i^l (x_t) is represented as;

##EQU13## wherein M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

5. A method for performing acoustic model estimation in order to optimize classification accuracy on feature vectors derived from a speaker with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond, the method comprising the steps of:

(a) initializing an acoustic model for each phone;

(b) evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from the phone and a second component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from other phones;

(c) adapting the acoustic model for selected phones so as to one of increase the first component of the discriminant measure for the phone and decrease the second component of the discriminant measure for the phone, the adapting step yielding a new acoustic model for each selected phone;

(d) evaluating the merit of the new acoustic models for each phone adapted in step (c) utilizing the two component discriminant measure;

(e) comparing results of the evaluation performed in step (b) with results of the evaluation of step (d) for each phone, and if one of the first component of the discriminant measure has increased and the second component of the discriminant measure has decreased, then the new acoustic model is kept for that phone, else the acoustic model originally initialized is kept;

(f) estimating parameters associated with each acoustic model kept for each phone in order to substantially optimize the objective function; and

(g) evaluating termination criterion to determine if the parameters of the acoustic models are substantially optimized.

6. The method of claim 5, further comprising the step of repeating steps (b) through (g) if the termination criterion has not been substantially satisfied.

7. The method of claim 5, wherein the first component of the two component discriminant measure is represented as:

##EQU14## where x_t represents the feature vectors and T₁ represents a normalizing factor and P_c^l (x_t) is represented as;

##EQU15## where M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

8. The method of claim 5, wherein the second component of the two component discriminant measure is represented as:

##EQU16## where x_t represents the feature vectors and T₂ represents a normalizing factor and where P_i^l (x_t) is represented as;

##EQU17## where M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

9. The method of claim 5, wherein the first component of the two component discriminant measure is represented as:

##EQU18## where x_t represents the feature vectors and T₁ represents a normalizing factor and where P_c^l (x_t) is represented as;

##EQU19## where M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

10. The method of claim 5, wherein the second component of the two component discriminant measure is represented as:

##EQU20## where x_t represents the feature vectors and T₂ represents a normalizing factor and where P_i^l (x_t) is represented as;

##EQU21## where M_l represents the acoustic model for phone l, M_j represents the acoustic model for phone j, C(x_t) represents a correct phone and F(x_t) represents confusable phones.

11. The method of claim 5, wherein the adapting step further includes comparing the first component to a threshold value to determine whether the acoustic model of a phone is to be adapted.

12. The method of claim 5, wherein the adapting step further includes comparing the second component to a threshold value to determine whether the acoustic model of a phone is to be adapted.

13. The method of claim 5, wherein the adapting step further includes comparing a ratio of the first component to the second component to a threshold value to determine whether the acoustic model of a phone is to adapted.

14. The method of claim 5, wherein step (a) further includes selecting an acoustic model type, an acoustic model complexity and initial acoustic model parameters.

15. The method of claim 5, wherein the acoustic models are categorized as gaussian mixtures.

16. The method of claim 15, wherein step (a) further includes selecting a number of mixture components and means, variances and priors distributions of the mixture components.

17. The method of claim 16, wherein step (c) further includes one of increasing and decreasing the number of mixture components of the acoustic model depending on a comparison to at least one threshold value associated with the first and second components of the discriminant measure.

18. The method of claim 16, wherein the new acoustic model is kept if after increasing the number of mixture components the first component of the discriminant measure increases more than the second component of the discriminant measure.

19. The method of claim 16, wherein the new acoustic model is kept if after decreasing the number of mixture components the second component of the discriminant measure decreases more than the first component of the discriminant measure.

20. The method of claim 16, wherein step (f) further includes substantially optimizing the means, variances and priors distributions of the mixture components of the kept acoustic model.

21. The method of claim 16, wherein step (g) further includes comparing a median value of the first component of the discriminant measure to a termination threshold value and if the median value is greater than the termination threshold value then the parameters are considered to be substantially optimized, else repeat steps (b) through (g).

22. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for performing acoustic model estimation in order to optimize classification accuracy on feature vectors derived from a speaker with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond, the method comprising the steps of:

(a) initializing an acoustic model for each phone;

(b) evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from the phone and a second component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from other phones;

(c) adapting the acoustic model for selected phones so as to one of increase the first component of the discriminant measure for the phone and decrease the second component of the discriminant measure for the phone, the adapting step yielding a new acoustic model for each selected phone;

(d) evaluating the merit of the new acoustic models for each phone adapted in step (c) utilizing the two component discriminant measure;

(e) comparing results of the evaluation performed in step (b) with results of the evaluation of step (d) for each phone, and if one of the first component of the discriminant measure has increased and the second component of the discriminant measure has decreased, when the new acoustic model is kept for that phone, else the acoustic model originally initialized is kept;

(f) estimating parameters associated with each acoustic model kept for each phone in order to substantially optimize the objective function; and

(g) evaluating termination criterion to determine if the parameters of the acoustic models are substantially optimized.

23. The program storage device of claim 22, further comprising the step of repeating steps (b) through (g) if the termination criterion has not been substantially satisfied.