Speech classification apparatus, speech classification method, and speech classification program

• US 8,630,853 B2
• Filed: 03/13/2008
• Issued: 01/14/2014
• Est. Priority Date: 03/26/2007
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Koshinaka, Takafumi
• Primary Examiner(s): AZAD, ABUL K
• Application Number:
  US12/593,323
• Publication Number:
  US 20100138223A1
• Time in Patent Office: 2,133 Days
• Field of Search: 704240-245
• US Class Current: 704/245
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

A speech classification apparatus includes a speech classification probability calculation unit that calculates a probability (probability of classification into each cluster) that a latest one of the speech signals (speech data) belongs to each cluster based on a generative model which is a probability model, and a parameter updating unit that successively estimates parameters that define the generative model based on the probability of classification of the speech data into each cluster calculated by the speech classification probability calculation unit.

Claims

1. A speech classification apparatus that classifies speech signals into clusters based on vocal similarity, comprising:

a speech classification probability calculation unit that calculates a probability that a latest input one of the speech signals sequentially input belongs to each of the clusters, based on a probability model for probabilistically determining to which cluster a certain speech signal belongs; and

a parameter updating unit that successively estimates values of parameters that define the probability model using each probability calculated by the speech classification probability calculation unit and a sufficient statistic calculated based on a speech signal inputted before the latest input one of the speech signals;

the speech classification probability calculation unit calculating each probability based on the probability model

defined by latest values of the parameters successively estimated by the parameter updating unit.

2. The speech classification apparatus according to claim 1, comprising:

a speech classification probability updating unit that recalculates probabilities that the speech signals which have been input within predetermined times in the past belong to the respective clusters, based on the probability model defined by the latest values of the parameters successively estimated by the parameter updating unit;

the parameter updating unit estimating the parameters that define the probability model using each of the probabilities calculated by the speech classification probability updating unit and a sufficient statistic calculated based on a speech signal inputted before the predetermined times.

3. The speech classification apparatus according to claim 1, comprising:

a new cluster registration unit that generates a probability model that defines a new cluster to which the input speech signal belongs, assuming a case where the input speech signal does not belong to any cluster; and

a cluster number determination unit that determines whether or not to add the new cluster from a result of parameter estimation by the parameter updating unit based on a result of calculation using the probability model generated by the new cluster registration unit.

4. The speech classification apparatus according to claim 1, wherein the probability model includes a hidden Markov model in which states are in a one-to-one correspondence with the clusters.

5. The speech classification apparatus according to claim 4, wherein the probability model includes a hidden Markov model associated with a Gaussian mixture distribution having a number of mixtures corresponding to a number of types of phonemes.

6. The speech classification apparatus according to claim 2, comprising:

an update target speech selection unit that determinants whether or not to cause the speech classification probability updating unit to recalculate a probability that each of the speech signals input within the predetermined times in the past belongs to each cluster.

7. The speech classification apparatus according to claim 6, wherein the update target speech selection unit determines whether or not recalculation of the probability that each of the speech signals belongs to each cluster is needed, based on an entropy of the calculated probability that each of the speech signals belongs to each cluster at a time of the determination as to the recalculation.

8. The speech classification apparatus according to claim 3, wherein in case the speech signal where the cluster to which the speech signal should belong is known is provided in advance, the new cluster registration unit generates a probability model that defines the cluster to which the speech signal should belong.

9. A speech classification method that classifies speech signals into clusters based on vocal similarity, comprising:

calculating a probability that a latest input one of the speech signals sequentially input belongs to each cluster, based on a probability model for probabilistically determining to which cluster a certain speech signal belongs;

successively estimating parameters that define the probability model using the probability and a sufficient statistic calculated based on a speech signal inputted before the latest input one of the speech signals; and

calculating a probability that at least next input speech signal belongs to each cluster, based on the probability model defined by the successively estimated parameters.

10. The speech classification method according to claim 9, comprising:

recalculating probabilities that the speech signals which have been input within predetermined times in the past belong to the respective clusters, based on the probability model defined by the latest values of the successively estimated parameters; and

estimating the parameters that define the probability model using each of the recalculated probabilities and a sufficient statistic calculated based on a speech signal inputted before the predetermined times.

11. The speech classification method according to claim 9 comprising:

generating a probability model that defines a new cluster to which the input speech signal belongs, assuming a case where the input speech signal does not belong to any cluster; and

determining whether or not to add the new cluster from a result of parameter estimation based on a result of calculation using the generated probability model.

12. The speech classification method according to claim 9, wherein the probability model includes a hidden Markov model in which states are in a one-to-one correspondence with the clusters.

13. The speech classification method according to claim 12, wherein the probability model includes a hidden Markov model associated with a Gaussian mixture distribution having a number of mixtures corresponding to a number of types of phonemes.

14. The speech classification method according to claim 10, comprising:

determining whether or not to need recalculation of a probability that each of the speech signals input within the predetermined times in the past belongs to each cluster.

15. The speech classification method according to claim 14, comprising:

determining whether or not to need recalculation of the probability that each of the speech signals belongs to each cluster, based on an entropy of the calculated probability that each of the speech signals belongs to each cluster at a time of the determination as to the recalculation.

16. The speech classification method according to claim 10, comprising:

Generating a probability model that defines the cluster to which the speech signal should belong when the speech signal where the cluster to which the speech data should belong is known is provided in advance.

17. A non-transitory computer-readable recording medium storing a speech classification program that classifies speech signals into clusters based on vocal similarity, the speech classification program causing a computer to execute:

a probability calculation processing that calculates a probability that a latest input one of the speech signals sequentially input belongs to each cluster, based on a probability model for probabilistically determining to which cluster a certain speech signal belongs; and

a parameter update processing that successively estimates parameters that define the probability model using each probability calculated by the speech classification probability calculation processing and a sufficient statistic calculated based on a speech signal inputted before the latest input one of the speech signals;

the probability calculation processing calculating each probability based on the probability model defined by latest values of the successively estimated parameters.

18. The non-transitory computer-readable recording medium according to claim 17, wherein the speech classification program causes the computer to execute:

a probability recalculation processing that recalculates probabilities that the speech signals which have been input within predetermined times in the past belong to the respective clusters, based on the probability model defined by the latest values of the successively estimated parameters,

the parameter updating processing successively estimating the parameters that define the probability model using each of the probabilities calculated by the probability recalculation process and a sufficient statistic calculated based on a speech signal inputted before the predetermined times.

19. The non-transitory computer-readable recording medium according to claim 17, wherein the speech classification program causes the computer to execute the processing comprising:

generating a probability model that defines a new cluster to which the input speech signal belongs, assuming a case where the input speech signal does not belong to any cluster; and

determining whether or not to add the new cluster from a result of parameter estimation based on a result of calculation using the probability model that defines the new cluster.

20. The non-transitory computer-readable recording medium according to claim 17, wherein the probability model includes a hidden Markov model in which states are in a one-to-one correspondence with the clusters.

21. The non-transitory computer-readable recording medium according to claim 20, wherein the probability model includes a hidden Markov model associated with a Gaussian mixture distribution having a number of mixtures corresponding to a number of types of phonemes.

22. The non-transitory computer-readable recording medium according to 18, wherein the speech classification program causes the computer to execute the processing comprising:

determining whether or not to need recalculation of a probability that each of the speech signals input within the predetermined times in the past belongs to each cluster.

23. The non-transitory computer-readable recording medium according to claim 22, wherein the speech classification program causes the computer to execute the processing comprising:

determining whether or not recalculation of the probability that each of the speech signals belongs to each cluster is needed in the determination process, based on an entropy of the calculated probability that each of the speech signals belongs to each cluster at a time of the determination as to the recalculation.

24. The non-transitory computer-readable recording medium according to claim 17, wherein the speech classification program causes the computer to execute the processing comprising:

generating a probability model that defines the cluster to which the speech signal should belong when the speech signal where the cluster to which the speech signal should belong is known is provided in advance.

25. A speech classification system that performs a clustering process which generates a cluster in response to each of sequentially input speech data on-line, the system comprising:

a speech classification probability calculation unit that derives a probability that a latest one of the sequentially input speech data belongs to each cluster, using a generation model which is defined by parameter values stored in parameter storage unit, assuming a speech data distribution, and storing the probability in speech classification probability storage unit;

an update target speech selection unit that determines whether or not recalculation of the probability that each of the speech data belongs to each cluster is needed according to a magnitude relation between a predetermined threshold value and an indicator, the indicator being obtained by reversing a sign of an entropy of the probability that each of the speech data belongs to each cluster;

a speech classification probability updating unit that derives a probability that speech data, the probability of which is determined to be needed by the update target speech selection unit, out of predetermined items of the sequentially input speech data except the latest speech data belongs to each cluster and that updates the speech classification probability storage unit; and

a parameter updating unit that calculates sufficient statistics necessary for calculating the generation model on each of the numbers of clusters and uses each of the probabilities calculated by the speech classification probability calculation unit and a sufficient statistic calculated based on a speech data inputted before the latest one of the sequentially input speech data to estimate parameter values of the generation model, assuming a current number of clusters and some numbers of clusters in the vicinity of the current number of clusters, based on results of calculations by the speech classification probability calculation unit and the speech classification probability updating unit, and that successively updates the parameter values in the parameter storage unit with the estimated parameter values.

26. The speech classification system according to claim 25, comprising:

a new speaker registration unit that reads parameters and sufficient statistics of the generation model stored in the parameter storage unit and generating a generation model with the number of clusters being incremented by one; and

a cluster number determination unit that determines an optimal number of clusters among the some numbers of clusters assumed by the parameter updating unit from a result of estimation of the parameter values of the generation model by the parameter updating unit, and storing sufficient statistics and parameter values corresponding to the determined number of clusters in the parameter storage unit.