Adaptation of Compound Gaussian Mixture models

1. A method of enhancing speech recognition in noisy environments, said method comprising the steps of:

providing at least one initial Compound Gaussian Mixture model comprising a formally equivalent model structured into parameters comprising atoms, wherein Gaussian components are sliced into Gaussian components of smaller dimension termed bands, said bands being clustered into atoms;

applying an adaptation algorithm to compute an adapted set of parameters to be utilized in an adapted Compound Gaussian Mixture model using enrollment data of a speaker;

wherein said adaptation algorithm comprises:

computing an adapted set of parameters given enrollment data of a speaker via linear combination of (1) an initial set of parameters and (2) an estimated set of parameters, said estimated set estimated via calculating a maximum likelihood estimate according to a maximum likelihood criterion on the enrollment data;

wherein the adaptation algorithm involves reducing mismatch between characteristics of speech data that are specific to a speaker in an acoustic environment and characteristics of the initial Compound Gaussian Mixture model trained on general speech data; and

wherein a processor is used to construct the adapted Compound Gaussian Mixture model to enhance speech recognition in the noisy environment.

2. The method according to claim 1 , wherein said adaptation algorithm allows an adaptation of a mean and a variance of the atoms and does not constrain atoms of a same band to undergo a same transform.

3. The method according to claim 2 , wherein the adaptation algorithm further comprises at least one combination weight.

4. The method according to claim 3 , wherein said combination weights are generally inversely proportional to a confidence measure associated with the estimated set of parameters.

5. The method according to claim 2 , wherein said adaptation algorithm further comprises performing an expectation-maximization procedure in yielding the estimated set of parameters.

6. The method according to claim 1 , further comprising the steps of:

obtaining speech enrollment data; and

performing a feature extraction on the speech enrollment data to yield feature vectors, wherein said feature vectors are used in computing the adapted set of parameters.

7. The method according to claim 1 , further comprising the steps of:

obtaining speech enrollment data; and

adding noise for speech enrollment data collected in at least one non-noisy environment.

8. An apparatus for enhancing speech recognition in noisy environments, said apparatus comprising:

an arrangement for providing at least one initial Compound Gaussian Mixture model comprising a formally equivalent model structured into parameters comprising atoms, wherein Gaussian components are sliced into Gaussian components of smaller dimension termed bands, said bands being clustered into atoms;

an arrangement for applying an adaptation algorithm to compute an adapted set of parameters to be utilized in an adapted Compound Gaussian Mixture model using enrollment data of a speaker;

wherein said adaptation algorithm comprises:

computing an adapted set of parameters given enrollment data of a speaker via linear combination of (1) an initial set of parameters and (2) an estimated set of parameters, said estimated set estimated via calculating a maximum likelihood estimate according to a maximum likelihood criterion on the enrollment data;

wherein the adaptation algorithm involves reducing mismatch between characteristics of speech data that are specific to a speaker in an acoustic environment and characteristics of the initial Compound Gaussian Mixture model trained on general speech data; and

wherein a processor is used to construct the adapted Compound Gaussian Mixture model to enhance speech recognition in the noisy environment.

9. The apparatus according to claim 8 , wherein the adaptation algorithm allows an adaptation of a mean and a variance of the atoms and does not constrain atoms of a same band to undergo a same transform.

10. The apparatus according to claim 9 , wherein the adaptation algorithm further comprises at least one combination weight.

11. The apparatus according to claim 10 , wherein said combination weights are generally inversely proportional to a confidence measure associated with the estimated set of parameters.

12. The apparatus according to claim 9 , wherein said arrangement for applying an adaptation algorithm is adapted to perform an expectation-maximization procedure in yielding the estimated set of parameters.

13. The apparatus according to claim 8 , further comprising:

an arrangement for obtaining speech enrollment data; and

an arrangement for performing a feature extraction on the speech enrollment data to yield feature vectors, wherein said feature vectors are used in computing the adapted set of parameters.

14. The apparatus according to claim 8 , further comprising:

an arrangement for obtaining speech enrollment data; and

an arrangement for adding noise for speech enrollment data collected in at least one non-noisy environment.

15. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for enhancing speech recognition in noisy environments, said method comprising the steps of:

providing at least one initial Compound Gaussian Mixture model comprising a formally equivalent model structured into parameters comprising atoms, wherein Gaussian components are sliced into Gaussian components of smaller dimension termed bands, said bands being clustered into atoms;

applying an adaptation algorithm to compute an adapted set of parameters to be utilized in an adapted Compound Gaussian Mixture model using enrollment data of a speaker;

wherein said adaptation algorithm comprises:

computing an adapted set of parameters given enrollment data of a speaker via linear combination of (1) an initial set of parameters and (2) an estimated set of parameters, said estimated set estimated via calculating a maximum likelihood estimate according to a maximum likelihood criterion on the enrollment data;

wherein the adaptation algorithm involves reducing mismatch between characteristics of speech data that are specific to a speaker in an acoustic environment and characteristics of the initial Compound Gaussian Mixture model trained on general speech data; and

wherein a processor is used to construct the adapted Compound Gaussian Mixture model to enhance speech recognition in the noisy environment.