Calculating cost measures between HMM acoustic models

• US 20080059184A1
• Filed: 08/22/2006
• Published: 03/06/2008
• Est. Priority Date: 08/22/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: K. Soong, Frank Kao-Ping, Zhou, Jian-Lai, Liu, Peng
• Granted Patent: US 8,234,116 B2
• Time in Patent Office: Days
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

Measurement of Kullback-Leibler Divergence (KLD) between hidden Markov models (HMM) of acoustic units utilizes an unscented transform to approximate KLD between Gaussian mixtures. Dynamic programming equalizes the number of states between HMMs having a different number of states, while the total KLD of the HMMs is obtained by summing individual KLDs calculated by state pair by state pair comparisons.

Claims

1. A method for measuring the total Kullback-Leibler Divergence (KLD) of two hidden Markov models (HMM):

calculating an individual KLD for each pair of states, state by state, for the two HMMs; and

summing the individual KLDs together to obtain a total KLD for the two HMMs; and

outputting the total KLD.

2. The method of claim 1 and further comprising prior to the step of calculating, modifying at least one of the HMMs to equalize the number of states.

3. The method of claim 2 wherein the step of modifying includes assigning an associated penalty value to the modification.

4. The method of claim 3 wherein the step of modifying includes modifying at least one of the HMMs in a manner to minimize the penalty value.

5. The method of claim 2 wherein modifying includes inserting a state in one of the HMM models.

6. The method of claim 2 wherein modifying includes deleting a state in one of the HMM models.

7. The method of claim 2 wherein modifying includes substituting a state for one of the states in one of the HMM models.

8. The method of claim 2 wherein modifying includes inserting a non-skippable short pause (sp) state in one of the HMM models.

9. The method of claim of claim 2 wherein modifying includes performing a series of operations to at least one of the HMMs, the operations taken from a set of operations including inserting a state, deleting a state and substituting a state, wherein each operation includes a corresponding penalty value, and wherein modifying further comprises modifying at least one of the HMMs in a manner to minimize a total of the penalty values.

10. A computer readable medium having instructions which when executed by a computer perform a method for measuring the total Kullback-Leibler Divergence (KLD) of two hidden Markov models (HMM), the instructions comprising:

modifying at least one of the HMMs to equalize the number of states;

calculating an individual KLD for each pair of states, state by state, for the two HMMs; and

summing the individual KLDs together to obtain a total KLD for the two HMMs.

11. The computer readable medium of claim 10 wherein modifying includes performing a series of operations to at least one of the HMMs, the operations taken from a set of operations including inserting a state, deleting a state and substituting a state, wherein each operation includes a corresponding penalty value, and wherein modifying further comprises modifying at least one of the HMMs in a manner to minimize a total of the penalty values.

12. The computer readable medium of claim 10 wherein the step of modifying includes assigning an associated penalty value to the modification.

13. The computer readable medium of claim 12 wherein the step of modifying includes modifying at least one of the HMMs in a manner to minimize the penalty value.

14. The computer readable medium of claim 13 wherein modifying includes inserting a state in one of the HMM models.

15. The computer readable medium of claim 13 wherein modifying includes deleting a state in one of the HMM models.

16. The computer readable medium of claim 13 wherein modifying includes substituting a state for one of the states in one of the HMM models.

17. The computer readable medium of claim 13 wherein modifying includes inserting a non-skippable short pause (sp) state in one of the HMM models.

18. A system for comparing a pair of acoustic HMM models comprises a KLD (Kullback-Leibler Divergence) calculating module configured to receive the pair of acoustic models and calculate the KLD between the pair of acoustic models based on modifying at least one of the acoustic hidden Markov models (HMM) models to equalize the number of states and calculating the KLD from individual KLDs for each pair of states, state by state, for the two acoustic HMM models, the KLD calculating module further configured to provide as an output as least one of the calculated KLD and information related to sorting of the acoustic HMM models.

19. The system of claim 18 and further comprising an application processing module configured to receive the output and perform a task related to processing speech units.

20. The system of claim 18 wherein the KLD calculating module is configured to perform a series of operations to at least one of the acoustic HMM models, the operations taken from a set of operations including inserting a state, deleting a state and substituting a state, wherein each operation includes a corresponding penalty value, and wherein KLD calculating module is configured to modify at least one of the acoustic HMM models in a manner to minimize a total of the penalty values.