Sound processing registering a word in a dictionary

• US 7,228,276 B2
• Filed: 04/01/2002
• Issued: 06/05/2007
• Est. Priority Date: 03/30/2001
• Status: Expired due to Fees

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Lucke, Helmut, Omote, Masanori
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US10/296,797
• Publication Number:
  US 20040030552A1
• Time in Patent Office: 1,891 Days
• Field of Search: 704/243
• US Class Current: 704/243
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

The present invention relates to a voice recognition apparatus capable of easily registering a word which has not been registered. The registering of an unregistered word into a dictionary can be easily performed without causing a significant increase in the size of the dictionary. The clustering unit detects a cluster (detected cluster) to which a new unregistered word is to be added as a new member, from existing clusters obtained by clustering unregistered words. The unregistered word is added as a new member to the detected cluster, and the cluster is divided depending on the members of the cluster such that unregistered words which are acoustically similar to each other belong to the same cluster. The maintenance unit updates the word dictionary on the basis of the result of the clustering. The present invention may be applied to a robot including a voice recognition apparatus.

Claims

1. A voice recognition apparatus for processing an input voice and updating a dictionary used in a language processor in accordance with a result of the processing of the input voice, said voice recognition apparatus comprising:

cluster detection means for detecting, from existing clusters obtained by clustering voices, a cluster to which said input voice is to be added as a new member;

cluster division means for employing said input voice as the new member of the cluster detected by said cluster detection means and dividing said cluster depending on members of said cluster; and

update means for updating the dictionary on the basis of a result of division performed by said cluster division means

wherein the dictionary stores a phoneme sequence of a vocabulary to be recognized, and

wherein the update means updates the dictionary by adding, as a new entry to the dictionary, a phoneme sequence of a voice corresponding to a representative member representing members of a cluster created by the division or by replacing an entry of the dictionary with the phoneme sequence of the voice corresponding to the representative member representing members of the cluster created by the division.

2. A voice recognition apparatus according to claim 1, wherein said cluster division means divides the cluster by means of an EM (Expectation Maximum) method.

3. A voice recognition apparatus according to claim 1, wherein said cluster detection means calculates the score of the input voice with respect to each member of the cluster by determining the likelihood that the input voice is observed in the member of the cluster;

and selects, from the members of the cluster, a member giving a highest value to said score of the input voice and employs the selected member as a representative member representing the members of the cluster; and

determines the cluster having said representative member as a cluster to which the said input voice it to be added as a new member.

4. A voice recognition apparatus according to claim 1, wherein said input voice is an unregistered word which has not been registered, in advance, in the dictionary.

5. A voice recognition apparatus according to claim 3, wherein in a case in which, of members of the cluster, a member having a greatest sum of scores with respect to the other members of that cluster is employed as the representative member representing the members of that cluster, said cluster division means divides the cluster, to which the input voice has been added, into two clusters, that is, first and second clusters, such that two members of the original cluster become representative members of the first and second clusters, respectively.

6. A voice recognition apparatus according to claim 5, wherein in a case in which there are a plurality of combinations of two clusters consisting of first and second clusters, said cluster division means divides a cluster including the input voice as a member thereof into two clusters such that the cluster-to-cluster distance between the first cluster and the second cluster becomes smallest.

7. A voice recognition apparatus according to claim 6, wherein when a combination of two clusters consisting of first and second clusters is selected such that the cluster-to-cluster distance between the first cluster and the second cluster becomes smallest, if said smallest cluster-to-cluster distance is greater than a predetermined threshold, then said cluster division means divides the cluster including the input voice as a member thereof into said two clusters.

8. A voice recognition apparatus according to claim 5, further comprising storage means for storing the scores of the members of said cluster with respect to each member of each score.

9. A voice recognition apparatus according to claim 1, wherein the dictionary stores a phoneme sequence of a vocabulary to be recognized, and wherein said voice recognition apparatus further comprises voice recognition means for recognizing a voice on the basis of an acoustic model constructed in accordance with the phoneme sequence stored in said dictionary.

10. A voice recognition apparatus according to claim 9, wherein said acoustic model is an HMM (Hidden Markov Model).

11. A voice recognition apparatus according to claim 9, wherein said voice recognition means constructs an acoustic model corresponding to a phoneme sequence stored in the dictionary by concatenating HMMs in units of sub-words and recognizes a voice on the basis of said acoustic model.

12. A voice recognition apparatus according to claim 9, wherein said voice recognition means recognizes a voice also on the basis of a predetermined grammatical rule.

13. A voice recognition apparatus according to claim 12, wherein said voice recognition means extracts a particular period of the input voice in accordance with the predetermined grammatical rule;

and

said cluster detection means and said cluster division means perform their processes on said period of the input voice.

14. A voice recognition apparatus according to claim 13, wherein said voice recognition means extracts, as said particular period, a period of an unregistered word which is not registered in the dictionary from the input voice.

15. A voice recognition apparatus according to claim 14, wherein said voice recognition means extracts the period of the unregistered word in accordance with the predetermined grammatical rule using a garbage model.

16. A voice recognition apparatus according to claim 1, further comprising:

storage means for storing voice information associated with the input voice for use by the cluster detection means to detect a cluster; and

deletion means for, when it is determined that a specific condition is satisfied, deleting a particular one of pieces of voice information stored in the storage means.

17. A voice recognition apparatus according to claim 16, wherein the voice information stored in the storage means is digital data of the input voice.

18. A voice recognition apparatus according to claim 17, further comprising feature extraction means for extracting a feature vector indicating a specific feature of the input voice from the digital data of the input voice, wherein the voice information stored in the storage means is the feature vector of the input voice extracted by the feature extraction means.

19. A voice recognition apparatus according to claim 16, wherein said deletion means determines that the specific condition is satisfied when the number of members belonging to the specific cluster is greater than a predetermined value.

20. A voice recognition apparatus according to claim 16, further comprising not-referred-to time calculation means for calculating the not-referred-to time during which the cluster has not been referred to, wherein said deletion means determines that the specific condition is satisfied when the not-referred-to time of the cluster calculated by the not-referred-to time calculation means is greater than a predetermined value.

21. A voice recognition apparatus according to claim 16, further comprising input means for inputting a trigger signal, wherein said deletion means determines that the specific condition is satisfied when the trigger signal is input via the input means.

22. A voice recognition apparatus according to claim 16, further comprising emotion control means for controlling a parameter associated with emotion, wherein said deletion means determines that the specific condition is satisfied when the value of the parameter associated with emotion controlled by the emotion control means is greater than a predetermined value.

23. A voice recognition apparatus according to claim 16, further comprising used-memory-area calculation means for calculating the amount of used memory area of the storage means, wherein said deletion means determines that the specific condition is satisfied when the amount of used memory area calculated by the used-memory-area calculation means is greater than a predetermined value.

24. A voice recognition apparatus according to claim 16, further comprising clustering means for reclustering a voice corresponding to voice information stored in the storage means.

25. A voice recognition apparatus according to claim 24, wherein said update means updates the dictionary also on the basis of a result of reclustering performed by the clustering means.

26. A voice recognition apparatus according to claim 24, further comprising representative member selection means for, when the voice information is deleted by the deletion means, selecting a representative member representing the members of the cluster to which the voice corresponding to the deleted voice information belongs before being deleted, wherein if the new representative member selected by the representative member selection means is different from a previous representative member, said clustering means reclusters all voice information stored in the storage means.

27. A voice recognition apparatus according to claim 26, further comprising deletion cancel means for if the structure of the cluster reclustered by the clustering means is different from the immediately previous structure the cluster had before being reclustered by the clustering means, canceling the deletion process performed, by the deletion means, on the voice information such that the original state is obtained.

28. A voice recognition apparatus according to claim 26, wherein said clustering means performs reclustering by means of a k-means method.

29. A voice recognition method for processing an input voice and updating a dictionary used in a language processor in accordance with a result of the processing of the input voice, said voice recognition method comprising the steps of:

detecting, from existing clusters obtained by clustering voices, a cluster to which said input voice is to be added as a new member;

employing said input voice as the new member of the cluster detected in said cluster detection step and dividing said cluster depending on members of said cluster; and

updating the dictionary on the basis of a result of division performed in said cluster division step

wherein the dictionary stores a phoneme sequence of a vocabulary to be recognized, and

wherein the update means updates the dictionary by adding, as a new entry to the dictionary, a phoneme sequence of a voice corresponding to a representative member representing members of a cluster created by the division or by replacing an entry of the dictionary with the phoneme sequence of the voice corresponding to the representative member representing members of the cluster created by the division.

30. A computer-readable medium storing a computer program for causing a computer to perform voice processing for processing an input voice and updating a dictionary used in a language processor in accordance with a result of the processing of the input voice, said program comprising the steps of:

detecting, from existing clusters obtained by clustering voices, a cluster to which said input voice is to be added as a new member;

employing said input voice as the new member of the cluster detected in said cluster detection step and dividing said cluster depending on members of said cluster; and

updating the dictionary on the basis of a result of division performed in said cluster division step

wherein the dictionary stores a phoneme sequence of a vocabulary to be recognized, and

wherein the update means updates the dictionary by adding, as a new entry to the dictionary, a phoneme sequence of a voice corresponding to a representative member representing members of a cluster created by the division or by replacing an entry of the dictionary with the phoneme sequence of the voice corresponding to the representative member representing members of the cluster created by the division.