JPH0683384A - A device for automatic detection and identification of utterance intervals of multiple speakers in speech - Google Patents

Abstract

(57) [Summary] [Object] The main feature of the present invention is to provide an automatic detection and identification device for utterance intervals of a plurality of speakers in a voice capable of detecting and identifying voice intervals of an arbitrary number of unknown speakers. And [Structure] The input speech 1 is converted into a time series 3 of feature vectors by a speech feature extraction unit 2, and a code sequence 7 is generated by a common codebook 5 created by a common codebook creation unit 4 by a quantization unit 6.
The appearance frequency of each code is calculated for each voice section by the voice section start / end detection unit 8, the appearance probability calculation unit 10 creates a set 11 of appearance probabilities, and the cluster analysis unit 12 generates several clusters. It is divided into 13, and the class of the voice section is determined based on the information of the cluster 13.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic utterance section detection / identification apparatus for a plurality of speakers in a voice, and more particularly to an automatic utterance section automatic detection for automatically detecting and identifying unknown utterance sections of a plurality of speakers. The present invention relates to an identification device.

[0002]

2. Description of the Related Art Automatic detection and identification of multiple persons in speech, identification of multiple languages in speech, identification of non-speech, identification of speech and noise, acoustic language model When creating, it is necessary to identify the speaker's utterance section.

Conventionally, in order to detect the utterance section by a plurality of speakers, speaker registration is performed in advance by using the voices of the respective speakers, and the utterance section of the speaker is used by a speaker identification technique. The method of detecting and identifying is used. However, in order to detect and identify the speech section, it is necessary to register in advance, and it has been impossible to detect and identify the speech section of an arbitrary number of unknown speakers.

Therefore, a main object of the present invention is to provide an automatic utterance section detection / identification apparatus for a plurality of speakers in a voice, which can detect and identify the voice sections of an arbitrary number of unknown speakers.

[0005]

The invention according to claim 1 is
A voice feature extraction means for extracting a feature pattern from the input voice, a common codebook creation means for creating a common code,
Quantization means for quantizing the feature pattern extracted by the voice feature extraction means with the common code created by the common codebook creation means, and appearance probability calculation means for calculating the appearance probability of the common code for a plurality of voice intervals. A cluster analysis means for performing a cluster analysis of the calculated appearance probabilities, and an identification means for detecting the appearance probabilities belonging to each analyzed cluster and identifying the voice section corresponding to the appearance probabilities.

In the invention according to claim 2, the start and end of a plurality of voice sections are predetermined. The invention according to claim 3 is
The start and end of a plurality of voice sections are automatically detected.

The invention according to claim 4 automatically determines the number of speakers when the number of speakers is not given in advance.

The invention according to claim 5 calculates the appearance probability of the utterance section of a plurality of speakers in a voice having a noise cluster corresponding to a speaker-independent noise section.

According to a sixth aspect of the present invention, a voice feature extracting means for extracting a feature pattern from the input voice, a common codebook producing means for producing a common code, and a common codebook producing means for the extracted feature pattern. Quantization means that quantizes with the common code created by, update means that updates the code appearance probability, transition probability and initial state probability in the state of the Ergodic Hidden Markov Model based on a given initial value, and And a means for decoding speech using the obtained Ergodic hidden Markov model.

According to a seventh aspect of the present invention, a voice feature extracting means for extracting a feature pattern from an input voice and a voice feature quantity in a state of a mixed continuous distribution type ergodic hidden Markov model based on a preset initial value. Updating means for updating the occurrence probability, branching probability, transition probability and initial state probability of, the judging means for judging the update stop condition, and the obtained continuous continuous distribution type ergodic hidden Markov model for decoding speech It is configured with means.

[0011]

The automatic speech segment detection / identification device for a plurality of speakers in a voice according to the present invention extracts a characteristic pattern from an input voice, quantizes the extracted characteristic pattern with a common code, and divides it into a plurality of speech periods. On the other hand, the probability of appearance of the common code is calculated. A cluster analysis is performed on the set of appearance probabilities to divide the appearance probabilities into a plurality of clusters. At this time, if the number of speakers is given in advance, the division is performed until the specified number is reached in the cluster analysis, and if the number of speakers is unknown, the evaluation standard obtained by the cluster analysis satisfies a certain condition. At this time, the cluster division is stopped. It is determined that the appearance probabilities belonging to the respective clusters obtained here belong to the same speaker, and the voice section given the appearance probability is obtained from the speaker. Here, when a plurality of voice sections are not obtained in advance, the voice section can be automatically segmented using an automatic detection method. Further, the segmentation of the voice section and the speaker cluster identification of the section can be performed at the same time by using the ergodic hidden Markov model.

[0012]

1 is a block diagram of an embodiment of the present invention. With reference to FIG. 1, an input voice 1 is given to a voice feature extraction unit 2 and converted into a time series 3 of feature vectors. The common codebook creation unit 4 creates a common codebook 5 from the voice in advance or independently of it, and gives it to the quantization unit 6. The quantizing unit 6 converts the time series of feature vectors given from the voice feature extracting unit 2 into a code sequence 7. Sequence 7 of this code
Is given to the voice section start / end detection unit 8, and the voice section start / end detection unit 8 detects the start / end of the voice section and divides it into a plurality of voice sections. The set 9 of voice sections is given to the appearance probability calculation unit 10, the appearance frequency of each code is calculated for each voice section, and the set 11 of appearance probabilities is created and given to the cluster analysis unit 12.

The cluster analysis unit 12 divides the set of appearance probabilities into several clusters. If the number of clusters is designated in advance, it may be that number. On the other hand, when the number is not specified, the number is set according to the evaluation standard. As a means of cluster analysis, for example, a vector quantization method is used, and cluster analysis is performed with a threshold value for quantization distortion. Cluster analysis unit 12
The cluster 13 analyzed in step 1 is the voice section class discriminator 1
4, it is assumed that a speaker having the same appearance probability belonging to the cluster is uttered based on the information of the cluster, and a voice section corresponding to the appearance probability is detected as being uttered by the same speaker. Identified.

FIG. 2 is a block diagram of another embodiment of the present invention. The embodiment shown in FIG. 2 is the same as the embodiment shown in FIG. 1 except for the following points. That is, the voice section start / end detection unit 15 determines a designated voice category other than the speaker (for example, a plurality of language categories such as Japanese and English).
The beginning and end of the section corresponding to is detected and divided into a plurality of voice sections to create a set 9 of voice sections. Hereinafter, the appearance probability calculation unit 10 determines the appearance probability 11 as in the embodiment of FIG. It is calculated.

FIG. 3 is a block diagram of still another embodiment of the present invention. In FIG. 3, the speech feature extracting unit 2, the common codebook creating unit 4, and the quantizing unit 6 are the same as those in the embodiments of FIGS. 1 and 2, and the code string 7 converted by the quantizing unit 6 is a discrete ergodic. The parameter 17 is given to the HMM (Hidden Markov Model) calculation unit 16 and the parameter 17 is estimated. This parameter 17 is given to the voice back trace unit 18, and again based on the estimated parameter, the ergodic HMM.
Is used to calculate the optimum correspondence between the code string and the state of the voice, and the back trace information 19 is calculated. The back trace information 19 is given to the state corresponding unit 20 of the voice section, and the back trace information detects and identifies that the voice section belonging to each state is uttered by the same speaker.

FIG. 4 is a block diagram of another embodiment of the present invention. The embodiment shown in FIG. 4 uses the mixed continuous distribution type ergodic HMM calculation unit 3. The input voice 1 is converted into a time series 3 of feature vectors in the voice feature extraction unit 2, input to the mixed continuous distribution ergodic HMM calculation unit 23, and its parameter 24 is estimated. Based on the estimated parameters, the ergodic HMM is used again to calculate the optimum correspondence between the code string and the state by the voice back trace unit 6, and the back trace information 19 is calculated. This backtrace information 19
Is given to the state corresponding unit 20 of the voice section, and the back trace information 19 detects and identifies that the voice section belonging to each state is uttered by the same speaker. If the number of clusters is specified in advance, the number of this state can be used as the number. On the other hand, when the number is not specified, the number can be set according to the evaluation standard.
As one means, it is possible to use a threshold for the likelihood of HMM.

FIG. 5 is a block diagram of another embodiment of the present invention. Also in the embodiment shown in FIG. 5, the voice feature extraction unit 2, the common codebook creation unit 4, and the quantization unit 6 are similar to those in FIGS.
As in the embodiment of FIG. 3, the code string 7 converted by the quantizing unit 6 is supplied to the voice section and noise section start / end detecting unit 21. The voice section / noise section start / end detection unit 21 detects the start / end of a voice section and a noise section, divides the start section into a plurality of voice sections and noise sections, and creates a set 22 of voice and noise sections. The appearance probability calculation unit 10 calculates the appearance probability by calculating the appearance frequency of each code based on the set 22 of voice and noise intervals, and the appearance probability set 11
To the cluster analysis unit 12.

The cluster analysis unit 12 divides the appearance probability set 11 into several clusters. If the number of clusters is designated in advance, it can be set to that number. On the other hand, if the number is not designated, the number can be set according to the evaluation standard. As a means for cluster analysis, in the same way as the embodiment shown in FIG. 1 described above, when the vector quantization method is used, it is possible to use a threshold value for quantization distortion. Based on the information of the cluster 13, the cluster discriminating unit 14 for the voice section
It is assumed that the appearance probabilities belonging to No. 3 are uttered from the same speaker category and the noise category, and the speech and noise intervals corresponding to the appearance probabilities are detected and identified as those generated from the same category.

[0019]

As described above, according to the present invention, the characteristic pattern is extracted from the input voice, the characteristic pattern is quantized by the common code, and the appearance probability of the common code is calculated for each voice section. Then, by performing cluster analysis on the calculated appearance probabilities, by detecting the appearance probabilities belonging to each cluster and identifying the corresponding voice sections, it is possible to detect the voice sections of an arbitrary number of unknown speakers without registering them in advance. Can be identified.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a block diagram of another embodiment of the present invention.

FIG. 3 is a block diagram of still another embodiment of the present invention.

FIG. 4 is a block diagram of another embodiment of the present invention.

FIG. 5 is a block diagram of still another embodiment of the present invention.

[Explanation of symbols]

 1 Input Speech 2 Speech Feature Extraction Section 3 Feature Sequence 4 Common Codebook Creation Section 5 Common Codebook 6 Quantization Section 7 Code Sequence 8 and 15 Speech Section Start / End Detection Section 9 Speech Set 10 Occurrence Probability Calculation Section 11 Appearance Probability 12 cluster analysis unit 13 cluster 14 speech class determination unit 16 discrete ergodic HMM calculation unit 17 ergodic HMM parameter 18 speech backtrace unit 19 backtrace information 20 state correspondence unit for speech segment 21 speech segment and noise segment Edge detector 23 Mixed continuous distribution type ergodic HMM calculator

Claims (7)

[Claims] 1. A voice feature extracting means for extracting a feature pattern from input voice, a common codebook producing means for producing a common code, and a feature pattern extracted by the voice feature extracting means by the common codebook producing means. Quantization means for quantizing with the created common code, appearance probability calculation means for calculating the appearance probability of the common code with respect to a plurality of speech sections, and cluster analysis of the appearance probabilities calculated by the appearance probability calculation means A utterance section of a plurality of speakers in a voice, which includes a cluster analysis unit and an identification unit that detects an appearance probability belonging to each cluster analyzed by the cluster analysis unit and identifies a voice section corresponding to the appearance probability. Automatic detection and identification device. 2. The apparatus for automatically detecting and identifying utterance sections of a plurality of speakers in a voice according to claim 1, wherein the start and end points of the plurality of voice sections are predetermined. 3. The apparatus for automatically detecting and identifying utterance sections of a plurality of speakers in a voice according to claim 1, wherein the start and end of the plurality of voice sections are automatically detected. 4. The cluster analyzing means automatically determines the number of speakers when the number of speakers is not given in advance. Multi-speaker utterance section automatic detection and identification device. 5. The appearance probability calculation means calculates the appearance probability of a speech section of a plurality of speakers in a voice having a noise cluster corresponding to a speaker-independent noise section. Device for automatic detection and identification of utterance intervals of multiple speakers in a voice. 6. A voice feature extracting means for extracting a feature pattern from input voice, a common codebook producing means for producing a common code, and a feature pattern extracted by the voice feature extracting means by the common codebook producing means. Quantizing means for quantizing with the created common code, updating means for updating the appearance probability, transition probability and initial state probability of the code in the state of the ergodic hidden Markov model based on a given initial value, by the updating means A device for automatically detecting and identifying a utterance section of a plurality of speakers in a voice, comprising: a determining unit that determines a stop condition for updating; and a unit that decodes the voice using the obtained Ergodic hidden Markov model. 7. A voice feature extracting means for extracting a feature pattern from an input voice, the appearance probability of voice features in a state of a mixed continuous distribution ergodic hidden Markov model, a branch probability, based on an initial value given in advance. Update means for updating the transition probabilities and initial state probabilities, judging means for judging the update stop condition by the updating means, and means for decoding speech using the obtained continuous continuous distribution type ergodic hidden Markov model In addition, a device for automatic detection and identification of utterance intervals of multiple speakers in a voice.