Method and apparatus for training Hidden Markov Model

• US 5,890,114 A
• Filed: 02/28/1997
• Issued: 03/30/1999
• Est. Priority Date: 07/23/1996
• Status: Expired due to Term

• Current Assignee: Inphi Corporation
• Original Assignee: OKI Electric Industry Company Limited
• Inventors: Yi, Jie
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US08/808,256
• Time in Patent Office: 760 Days
• Field of Search: 704/251, 704/253, 704/254, 704/255, 704/256, 704/232, 704/245
• US Class Current: 704/256
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

HMM training method comprising a first parameter predicting step, a centroid state set calculating step, a reconstructing step, a second parameter predicting step and a control step. In the first parameter predicting step, a parameter of an HMM (hidden Markov model) is predicted based on training data. In the centroid state set calculating step, a centroid state set is calculated by clustering the state of said HMM whose parameter is predicted in the first parameter predicting step. In the reconstructing step, an HMM is reconstructed with using the centroid state calculated in the centroid state set calculating step. In the second parameter predicting step, predicted a parameter of the HMM reconstructed in the reconstructing step with using the training data. And, the centroid step is reexecuted by the control step in the case that a likelihood of the HMM whose parameter is predicted in the second parameter predicting step does not satisfy a predetermined condition.

Claims

1. A training method for generating parameters of an HMM (hidden Markov model) for speech recognition, comprising:

converting speech to digital sound signals;

producing training data from the digital sound signals;

obtaining an initial estimated parameter of the HMM based on the training data;

obtaining a set of centroid states, by

determining a state of the HMM using the initial estimated parameter, and

clustering the determined state of the HMM;

reconstructing the HMM with the set of centroid states;

obtaining a new estimated parameter of the reconstructed HMM; and

repeatedly reexecuting

the obtaining a set of centroid states,

the reconstructing the HMM, and

the obtaining a new estimated parameter,

until a variation value of likelihood of the new estimated parameter becomes not more than a preset value, wherein the reexecution of the obtaining a set of centroid states uses a last obtained new estimated parameter in place of the initial estimated parameter.

2. The training method as claimed in claim 1, wherein said converting speech to digital sound signals includes producing analog sound signals from the speech and producing the digital sound signals from the analog sound signals.

3. An HMM training method as claimed in claim 1 wherein:

said HMM is a triphone HMM.

4. An HMM training method as claimed in claim 1 wherein:

said HMM is a diphone HMM.

5. The training method as claimed in claim 1, wherein said converting speech to digital sound signals includes inputting the speech at a microphone connected to an A/D converter.

6. The training method as claimed in claim 5, wherein said producing training data from the digital sound signals includes producing sound feature parameters corresponding to features of the speech from which the digital sound signals are converted, and obtaining the training data by selecting a group of the produced sound feature parameters.

7. A training method for generating parameters of an HMM (hidden Markov model) for speech recognition, comprising:

converting speech to digital sound signals;

producing training data from the digital sound signals;

obtaining an initial estimated parameter of the HMM based on the training data;

obtaining a set of centroid states, by

determining a state of the HMM using the initial estimated parameter, and

clustering the determined state of the HMM;

performing reconstruction of the HMM, including

determining distances of the obtained centroid states from a state of interest,

selecting a plurality of the obtained centroid states based on the determined distances, and

reconstructing the HMM with the selected centroid states;

obtaining a new estimated parameter of the reconstructed HMM; and

repeatedly reexecuting

the obtaining a set of centroid states,

the performing reconstruction of the HMM, and

the obtaining a new estimated parameter,

until a variation value of likelihood of the new estimated parameter becomes not more than a preset value, wherein the reexecution of the obtaining a set of centroid states uses a last obtained new estimated parameter in place of the initial estimated parameter.

8. An HMM training method as claimed in claim 7 wherein said HMM is a diphone HMM.

9. An HMM training method as claimed in claim 7 wherein said HMM is a triphone HMM.

10. The training method as claimed in claim 7, wherein said converting speech to digital sound signals includes inputting the speech at a microphone connected to an A/D converter.

11. The training method as claimed in claim 7, wherein said converting speech to digital sound signals includes producing analog sound signals from the speech and producing the digital sound signals from the analog sound signals.

12. The training method as claimed in claim 11, wherein said producing training data from the digital sound signals includes producing sound feature parameters corresponding to features of the speech from which the digital sound signals are converted, and obtaining the training data by selecting a group of the produced sound feature parameters.

13. A training apparatus for generating parameters of an HMM (hidden Markov model) for speech recognition, comprising:

means for converting speech to digital sound signals;

means for producing training data from the digital sound signals;

means for obtaining an initial estimated parameter of the HMM based on the training data.

means for obtaining a set of centroid states, by

determining a state of the HMM using the initial estimated parameter, and

clustering the determined state of the HMM;

means for reconstructing the HMM with the set of centroid states;

means for obtaining a new estimated parameter of the reconstructed HMM; and

means for repeatedly controlling

the means for obtaining a set of centroid states based on a last obtained new estimated parameter,

the means for reconstructing the HMM, and

the means for obtaining of a new estimated parameter,

and determining that a variation value of likelihood of the last obtained new estimated parameter becomes not more than a preset value.

14. The training apparatus as claimed in claim 13, wherein said means for converting speech to digital sound signals includes means for producing analog sound signals from the speech and means for producing the digital sound signals from the analog sound signals.

15. The training method as claimed in claim 13, wherein said means for producing training data from the digital sound signals includes means for producing sound feature parameters corresponding to features of the speech from which the digital sound signals, and means for selecting a group of the produced sound feature parameters as the training data.

16. The training apparatus as claimed in claim 13, wherein said means for converting speech to digital sound signals includes a microphone and an A/D converter connected to the microphone.

17. A training apparatus for generating parameters of an HMM (hidden Markov model) for speech recognition, comprising:

means for converting speech to digital sound signals;

means for producing training data from the digital sound signals;

means for obtaining an initial estimated parameter of the HMM based on the training data;

means for obtaining a set of centroid states, by

determining a state of the HMM using the initial estimated parameter, and

clustering the determined state of the HMM;

means for performing reconstruction of the HMM, including

determining distances of the obtained centroid states from a state of interest,

selecting a plurality of the obtained centroid states based on the determined distances, and

reconstructing the HMM with the selected centroid states;

means for obtaining a new estimated parameter of the reconstructed HMM; and

means for repeatedly controlling

the means for obtaining of a set of centroid states based on a last obtained new estimated parameter,

the means for performing of reconstruction of the HMM, and

the means for obtaining of a new estimated parameter,

and determining that a variation value of likelihood of the last obtained new estimated parameter becomes not more than a preset value.

18. The training apparatus as claimed in claim 17, wherein said means for converting speech to digital sound signals includes means for producing analog sound signals from the speech and means for producing the digital sound signals from the analog sound signals.

19. The training method as claimed in claim 18, wherein said means for producing training data from the digital sound signals includes means for producing sound feature parameters corresponding to features of the speech from which the digital sound signals, and means for selecting a group of the produced sound feature parameters as the training data.

20. The training apparatus as claimed in claim 17, wherein said means for converting speech to digital sound signals includes a microphone and an A/D converter connected to the microphone.