JP2000122693A - Speaker recognizing method and speaker recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To shorten time required for recognizing a speaker while holding high recognizability by making a speaker model having a housing structure of a wood structure on the basis of a characteristic quantity extracted from an uttering sound of the speaker. SOLUTION: A speaker model of a single speaker is expressed by a wood structure. The speaker model making part 3 makes the speaker model by using a characteristic vector group calculated from voice data of the speaker inputted at learning time. In a method of using a code book of a characteristic vector as the speaker model, a representative vector is decided by one on respective clusters by clustering the characteristic vector of the speaker by any method. Aggregation of these representative vectors is used as the speaker model. The checkup part 5 checks up the characteristic vector of a recognizing object speaker calculated by the characteristic quantity calculating part 2 with respective speaker models stored in the speaker model storage part 4 to calculate a true degree or a distance to identify it as the person himself when the true degree exceeds a prescribed value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a speaker recognition method for identifying an individual using voice and a speaker recognition device having the same.

[0002]

2. Description of the Related Art Heretofore, identification of individuals at financial institutions or government offices has been performed using a registered seal or password. However, in such a conventional method, at present, accurate identification of an individual cannot be performed due to loss or theft of a seal, forgetting or leaking a personal identification number. Therefore, a method of performing individual identification using a person's voice has been proposed.

[0003] Speaker recognition methods for identifying individuals using voice include a content-dependent method of generating and recognizing specific words and a content-independent method of performing authentication for arbitrary occurrences. Separated. Generally, a higher recognition rate can be obtained by using the utterance content-dependent method, but a constraint that requires utterance of a specific word is added. The utterance content independent method also has an advantage that the longer the utterance length is, the higher the recognition rate can be obtained. The utterance content independent method that does not need to memorize a specific word and has a small burden on the user can be applied to various fields.

[0004] An example of a conventional speaker recognition method using a conventional utterance content independent method will be described below. At the time of learning, a feature amount (feature vector) is calculated from the collected voice data of a plurality of speakers, and then a speaker model is created for each speaker and stored. At the time of recognition, speech data of a speaker to be recognized is converted into a feature amount, and the converted feature amount is compared with a plurality of speaker models stored before to calculate likelihood. If the degree exceeds a certain value, the person is identified.

When speech data is converted into a feature value during learning, for example, a method of extracting a section of about 16 ms to 40 ms from the speech data every 8 ms to 16 ms and calculating a feature value for each section is known. Commonly used.

At the time of recognition, a feature amount is similarly calculated for each section, and the calculated feature amount is individually collated with a speaker model, and the collation results are integrated to obtain a final recognition result. Is calculated.

To create a speaker model, a Hidden Markov Model (hereinafter referred to as HM)
M), clustering the distribution of a plurality of feature vectors, holding representative feature vectors of each cluster, and using a codebook listing the held representative feature vectors. Are listed.

Further, in order to cope with a change in the surrounding environment, the likelihood is calculated, and the recognition result of the speaker model other than the speaker is used in addition to the recognition result of the speaker model of the speaker. Is generally used. For example, in addition to the speaker model of the subject, five other speaker models similar to the subject are calculated and the likelihood (distance) is calculated, and two values having a high likelihood (small distance) are used to determine the recognition result. A technique for performing normalization is also often used.

The method of creating a speaker model greatly affects the performance of speaker recognition. In speaker recognition, in order to avoid erroneous recognition, a speaker whose likelihood with the voice data of the subject is large (or the distance is small) and whose likelihood with the speech data of another person is small (or the distance is large) You need to create a model. Also, considering the burden on the user when performing speaker recognition, it is desirable to perform recognition with a minimum amount of voice data. However, if the utterance time is short, it becomes difficult to grasp the characteristics of the individual. For this reason, it is necessary to create a speaker model that maintains high recognition performance even for short utterances.

In order to maintain high recognition performance for short utterances, it is necessary to create a fine speaker model.
That is, in speaker recognition, for example, the distribution of a feature vector is obtained by combining several Gaussian distributions (mixture distribution).
When the HMM expressed by is used for the speaker model, the number of Gaussian distributions (distributions of the HMM) used for the mixture distribution is increased. When a codebook of feature vectors is used for a speaker model, a method of increasing the size of the codebook is used. However, in general, a calculation amount proportional to the number of distributions of the HMM and the size of the codebook is required. Therefore, the finer the model to be created, the longer the time required for recognition.

[0011]

As described above,
Creating a fine speaker model to improve the recognition performance for short utterances, on the other hand, increases the calculation amount of likelihood (distance) calculation for matching the input speech with the speaker model, There is a problem that the time required for speaker recognition becomes long.

Therefore, the present invention reduces the amount of calculation required when collating an input voice with a speaker model, and as a result,
It is an object of the present invention to provide a speaker recognition method and a speaker recognition device using the same that can reduce the time required for speaker recognition while keeping the fineness of a model to be created, that is, while maintaining high recognition performance. .

[0013]

(1) A speaker recognition method according to the present invention is a method for recognizing a speaker based on a speaker model created in advance from features extracted from uttered sounds of the speaker. In the speaker recognition method, a speaker model having a storage structure of a tree structure is created based on the feature amount extracted from the input utterance sound of the speaker and stored in the storage unit, and the speaker model of the input recognition target speaker is generated. The feature amount extracted from the uttered sound is compared with the stored speaker model to recognize the speaker to be recognized.

According to the present invention, even when a fine speaker model is created, the likelihood between the input speech and the speaker model is maintained while maintaining high recognition performance by forming the speaker model to be used in a tree structure. (Or distance) calculation amount can be reduced, and high-speed speaker recognition can be performed.

(2) The speaker recognition method according to the present invention is a speaker recognition method for recognizing a speaker based on a speaker model created in advance from feature amounts extracted from a speaker's utterance sound. A speaker model having a tree-structured storage structure is created based on the features extracted from the uttered voices of the selected speaker, stored in the storage unit, and extracted from the uttered voices of the input speaker to be recognized. To compare a feature amount with the stored plurality of speaker models and normalize a matching result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker. Is used to normalize the matching result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker using the selected speaker model. The speaker to be recognized is recognized.

According to the present invention, a speaker model to be used for normalization is determined by using a value obtained at the time of calculating a likelihood (or distance), so that the result of matching can be normalized. The accompanying increase in the amount of calculation can be suppressed.

(3) The speaker recognition apparatus of the present invention recognizes a speaker based on a speaker model created in advance from feature amounts extracted from the utterance of the speaker. Storage means for creating and storing a speaker model having a tree-structured storage structure based on the features extracted from the uttered sounds of the given speaker, and extracted from the uttered sounds of the input recognition target speaker And a recognition unit for recognizing the speaker to be recognized by comparing the feature amount with the speaker model stored in the storage unit.

According to the present invention, even if a fine speaker model is created, the likelihood between the input speech and the speaker model is maintained while maintaining a high recognition performance by forming a tree structure of the speaker model to be used. (Or distance) calculation amount can be reduced, and high-speed speaker recognition can be performed.

(4) The speaker recognition apparatus of the present invention recognizes a speaker based on a speaker model created in advance from features extracted from the utterance of the speaker. Storage means for creating and storing a speaker model having a tree-structured storage structure based on the features extracted from the uttered sounds of the given speaker, and extracted from the uttered sounds of the input recognition target speaker The feature amount is compared with the plurality of speaker models stored in the storage unit, and the matching result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker is normalized. Selecting means for selecting a speaker model to be used for conversion, and using the speaker model selected by the selecting means, the input features of the speaker to be recognized and the talk of the speaker to be recognized. Normalize the matching result with the speaker model, and Characterized by comprising a recognition means for identification, the.

According to the present invention, a speaker model to be used for normalization is determined by using a value obtained at the time of calculating a likelihood (or distance), so that a matching result can be normalized. The accompanying increase in the amount of calculation can be suppressed.

[0021]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows an example of the configuration of a speaker recognition apparatus (applying an independent speech content method) using a speaker recognition method according to the present embodiment. A feature amount calculating unit 2 that extracts a section from the extracted voice and calculates a feature amount (for example, a feature vector) for each extracted section; a speaker model creating unit 3 that creates a speaker model from the calculated feature amount; A speaker model storage unit 4 for storing the generated speaker model; a feature amount calculated from a voice input to the voice input unit 1 during speaker recognition; and a speaker model stored in the speaker model storage unit 4 (Dictionary) to calculate the likelihood (or the distance indicating the distance between the input speaker model and the speaker model as a dictionary),
If the calculated likelihood exceeds a predetermined value (or
(When the distance is smaller than a predetermined value) and a collating unit 5 for identifying the speaker as the subject.

An outline of the processing operation of the speaker recognition device having the configuration shown in FIG. 1 will be described with reference to a flowchart shown in FIG. At the time of learning (step S1), the voice input unit 1
After calculating a feature vector from the input speech data of a plurality of speakers in the feature amount calculation unit 2, a speaker model creation unit 3 creates a speaker model for each speaker, and a speaker model storage unit 4. (Steps S2 to S
5). At the time of recognition (step S1), a feature vector is calculated by the feature value calculation unit 2 from the voice data of the speaker to be recognized input from the voice input unit 1, and then the feature vector is calculated by the feature value calculation unit 2 in the collation unit 5. The feature vector of the speaker to be recognized is compared with each speaker model stored in the speaker model storage unit 4 to calculate likelihood or distance. For example, if the likelihood exceeds a predetermined value, the user is identified (step S6 to step S9).

FIG. 2 shows another example of the configuration of a speaker recognition apparatus (applying the utterance content independent method) using the speaker recognition method according to the present embodiment. The same parts as those in FIG. 1 are denoted by the same reference numerals, and only different parts will be described. That is, in FIG. 2, a normalization unit that normalizes the likelihood or distance between the feature amount of the speaker to be recognized calculated by the matching unit 5 and the speaker model of the speaker is added to the configuration shown in FIG. 6 is further added.

In step S8 in FIG. 3, the normalizing unit 6 calculates the likelihood or the distance in step S8 of FIG. The recognition result is normalized using the value of the recognition result for a speaker model other than the above. For example, in addition to the speaker model of the subject, the likelihood (or distance) is calculated for the other five speaker models similar to the subject, and the recognition result is calculated using two values with high likelihood (small distance). Is normalized. Based on the normalized recognition result (likelihood or distance), the speaker is identified as the person (when the likelihood exceeds a predetermined value or when the distance is smaller than the predetermined value).

1 and 2, a characteristic part of the present invention is a method of creating a speaker model performed by the speaker model creating section 3 in step S4 of FIG. That is, in the present invention, the tree structure shown in FIGS.
It is designed to represent a speaker model of one speaker. Hereinafter, this tree structure is called a tree structure type speaker model.

The speaker model creating section 3 creates a speaker model by using a feature vector group calculated from the speaker's voice data input at the time of learning. In the method using the codebook of the feature vector as a speaker model, the feature vector of the speaker is clustered by some method, and one representative vector is determined for each cluster. A set of these representative vectors is used as a speaker model. In the present embodiment, the codebook of this feature vector is used as a speaker model.

In step S4 of FIG.
Creates a tree-structured speaker model using feature vectors calculated from voice data of a plurality of speakers. FIG. 4 shows an example of the tree structure type speaker model. The tree structure type speaker model shown in FIG. 4 has a root node and terminal nodes (a1 to a3).
Although a two-layer tree structure having one stage of intermediate nodes (b1 to b8) is held between 2), a structure in which intermediate nodes are held in a plurality of stages can be used in principle. For example, as shown in FIG. 5, a certain terminal node (node c1 in FIG. 5,
A structure in which an intermediate node (nodes b2 and b3 in FIG. 5) is one stage for c2, c4, and c5) and an intermediate node is two stages for another terminal node (nodes a1 to a5 in FIG. 5). (In FIG. 5, the node c3 is the node a1 and a2, and the upper node b2 is the intermediate node. The nodes c3 and a5 are the nodes c3 and a5.
The upper node b3 may be an intermediate node).

Next, the procedure for constructing a tree-structured speaker model in the speaker model creation section 3 will be described with reference to the flowchart shown in FIG. In the tree-structured speaker model of the present invention, the feature vector is created by successively clustering the feature vector one more time than the number of intermediate nodes. The number of nodes in each stage is the number of clusters in each clustering, and one node is associated with one feature vector. A case where a tree structure type speaker model as shown in FIG. 4 is created will be described as an example. First, feature vectors are clustered into eight using a normal clustering method (step S1).
1). Here, since there are a plurality of feature vectors belonging to each cluster, representative vectors are obtained for each cluster (steps S12 to S13), and these are made to correspond to the nodes b1 to b8 (step S14).

Next, the second clustering is performed on the feature vectors belonging to each cluster, and the feature vectors belonging to each cluster are clustered again into four (step S15). Here, since each feature vector belonging to each cluster is one by one, each feature vector is made to correspond to each lower node (step S12, step S16).

In this embodiment, the terminal nodes (a1 to a
The final detailed feature vector is the intermediate node (8 nodes from b1 to b8) in up to 32 nodes (32 nodes).
, An intermediate representative vector is held, and a speaker model of one speaker consisting of a total of 40 representative vectors is constructed.

Conventionally, only 32 representative vectors corresponding to the terminal node are held, so that the size of the model is increased by the size of the intermediate node. Here, since the number of intermediate nodes was one, the speaker model was constructed by clustering twice. Even if the number of intermediate nodes increases, a model can be constructed by repeating such an operation until only one feature vector exists at the terminal node.

In the tree structure type speaker model, the upper node holds the average vector of the representative vectors of the terminal nodes connected to the lower node. That is, for example, in the tree structure type speaker model shown in FIG. 4, at the node b1, the average vector of the feature vectors held at each of the nodes a1 to a4 is held.

In the present invention, the distance between the input vector as a feature quantity extracted from the utterance sound of the speaker to be recognized and the speaker model is calculated according to the procedure shown in the flowchart of FIG.

In the present invention, the tree-structured speaker model expressing the result of clustering is traced from the root node to the terminal node to find the closest node, and the distance between the input vector and the vector of the node is determined. Is defined as the distance between the input vector and the speaker model, and when tracing the tree structure, the node currently being viewed is referred to as a node of interest. Here, a case of a tree structure type speaker model as shown in FIG. 4 will be described as an example.

The target node is set as the root node (step S21), and the distance between the input vector and the vectors of the nodes b1 to b8 connected to the target node is calculated (step S22). At this time, if, for example, b1 is the minimum distance, the target node is moved to the node b1 (step S23). Next, it is checked whether a lower node exists in the target node (step S24). If there is a lower node as in this case, step S22 is performed.
A, which is connected to the input vector and the node of interest
The distance from the vector of the node from 1 to a4 is calculated. The target node is moved to the node having the minimum distance based on the distance calculation (step S23). Again, it is checked whether a lower node exists in the target node (step S2).
4) If there is no lower node as in this case, the process proceeds to step S25, where the distance between the input vector and the vector of the node of interest is output as the distance between the input vector and the speaker model, and the process ends.

When 32 representative vectors are held as in the speaker model shown in FIG.
Using the conventional speaker model, the vector distance calculation is performed 32 times. On the other hand, in the present embodiment, a total of 12 vector distances are calculated for 8 nodes at the first stage and 4 nodes at the second stage. It can be seen that the calculation is completed and the distance calculation is speeded up by forming the speaker model into a tree structure.

Further, by making the speaker model into a tree structure,
The representative vector that should have been selected is not selected, and as a result, it is compared with the distance calculated by the conventional speaker model,
It is possible that larger distances may be calculated. If the error in the distance calculation affects the recognition accuracy, the error is changed by changing the node of interest to a node having up to s vectors counting from the smallest distance in step S23 in FIG. Can be smaller.

For example, in the case of the tree-structured speaker model shown in FIG. 4, when s = 2 in step S23 in FIG. As a result, a total of 16 distance calculations are required for 8 nodes at the first stage and 4 + 4 = 8 nodes at the second stage. Nevertheless, if it is compared with the conventional calculation amount of 32 times, it is sufficiently small, and if it is compared with the case where one minimum is selected, the error can be reduced.

Next, normalization of the calculated likelihood (normalization of distance) will be described. In general, in addition to the distance to the true speaker model, the distance to the other n speaker models is also calculated, and the distance is normalized using the following formula (1).

[0040]

(Equation 1)

In the equation (1), a distance d (X | Si) obtained from the input vector X and the speaker model Si is converted to a normalized distance using another speaker model Sj (1 ≦ j ≦ n). dnew
(X | Si). As the function f of the denominator, generally, a geometric mean or the like is often used.

For the n speaker models, a speaker with a voice similar to the subject, that is, a speaker model that outputs a small distance to the subject's utterance is specified in advance from the training speech. Usually, instead of using the distances for n speaker models, the calculation is often performed using k speaker models whose distance is small.

In the conventional method, the speaker normalization is n
It is necessary to calculate the distance from each speaker model. However, since the present invention uses a tree structure type speaker model,
It is possible to select k speaker models whose distances are approximately small and obtain the distances to those models without obtaining the distances to the n speaker models.

FIG. 8 is a flow chart for explaining the processing operation when normalizing the collation result between the feature vector extracted from the utterance sound of the speaker to be recognized and the speaker model of the speaker. A description will be given taking a tree structure type speaker model of FIG. 4 as an example.

Assuming that each root node of the n speaker models is the target node (step S31), the distance between the eight representative vectors of the nodes b1 to b8 in the first stage is first determined for each speaker model. The calculation is performed, and the minimum value among the eight is obtained (step S32). From the nodes (Node x ′) of the minimum value obtained from each of the n speaker models, t (> k) nodes having a small value are extracted (step S23), and they are newly added to the target node. (Step S33).

It is checked whether there is a node having a lower node among the t nodes newly set as the target node (step S34). In the tree structure type speaker model shown in FIG. 4, since there are nodes having lower nodes, the process proceeds to step S32, and only the t speaker models for which the new target node is set are set in the second stage. Is calculated.

In each of the t speaker models, t of the t distances determined in the second stage are taken in order from the smallest one (step S33). In the tree-structured speaker model shown in FIG. 4, there is no node that has a lower node (third stage node) this time, so the process proceeds to step S35, and from the t nodes, the nodes having smaller distances are sequentially arranged. The k nodes are extracted, and the distances are normalized using Equation (1) using speaker models each including the k nodes (step S35).

The speaker recognition in the case of performing the normalization is performed by using a tree-structured speaker model as shown in FIG.
Taking the case where k = 2 as an example, the effect will be described more specifically. In this case, in the conventional speaker model, it is necessary to calculate the distance for 10 speaker models consisting of 32 vectors, so that 320 vector distance calculations are required. On the other hand, according to the present invention, by structuring the speaker model, the calculation for each speaker model requires 8 × 10 = 80 times for the intermediate node in the first stage, The calculation for the terminal node is 4 × 10 = 40 times, total 1
Only 20 times.

Further, as shown in FIG.
By limiting at t = 5 speaker models at 33, 2
Since the calculation at the stage only requires the calculation at the second stage (calculation for four nodes) for each of the five speaker models,
4 × 5 = 20 times. If the normalization is performed according to the procedure shown in FIG. 8, in the case of the speaker recognition using the tree structure type speaker model, the vector distance calculation is performed 20 times compared with the case where the distances from all the speaker models are calculated. Can be omitted.

Further, by using the intermediate result at the intermediate node, which is output when obtaining the distance between the speaker model and the input vector, the speaker model which is apparently not used for distance normalization can be used. In other words, it is possible to speed up the likelihood normalization by terminating the calculation on the way and continuing to calculate only the speaker model that will eventually be required.

In the above embodiment, a tree-structure type speaker model as shown in FIG. 4 is created, but the present invention can be applied to any structure having a tree structure. In the present embodiment, the tree structure type speaker model is created by using the codebook of the feature vector as the speaker model.
A similar effect can be obtained by creating a tree-structured speaker model using the distribution of the HMM as the speaker model.

In speaker recognition, creating a fine speaker model in order to enhance recognition performance has the disadvantage of increasing the amount of calculation at the time of recognition, but as described above, as described above, According to, even if a fine speaker model is created, by using a tree structure for the speaker model to be used,
It is possible to reduce the amount of calculation at the time of calculating the likelihood (or distance) between the input speech and the speaker model while maintaining high recognition performance, thereby enabling high-speed speaker recognition.

When normalizing the collation result,
Since the likelihood (or distance) with the speaker model that is not calculated when the likelihood (or distance) is calculated is also required, the increase in the amount of calculation as compared with the case without normalization is conventionally It will be larger than that. However, according to the above embodiment, the speaker model to be used for the normalization is determined by using the value obtained in the calculation of the likelihood (or distance), so that the matching result can be normalized. The accompanying increase in the amount of calculation can be suppressed.

[0054]

As described above, according to the present invention,
The amount of calculation required for collating the input speech with the speaker model can be reduced, and the time required for speaker recognition can be reduced while maintaining high recognition performance.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of a speaker recognition device to which a speaker recognition method according to an embodiment of the present invention is applied.

FIG. 2 is a diagram showing another configuration example of the speaker recognition device to which the speaker recognition method according to the embodiment of the present invention is applied.

FIG. 3 is a flowchart schematically showing a processing operation of the speaker recognition device.

FIG. 4 is a diagram showing an example of a tree structure type speaker model.

FIG. 5 is a diagram showing another example of a tree structure type speaker model.

FIG. 6 is a flowchart showing a tree structure type speaker model construction processing procedure;

FIG. 7 is a flowchart illustrating a procedure for calculating a distance between an input vector as a feature amount extracted from a speech sound of a speaker to be recognized and a speaker model;

FIG. 8 is a flowchart for explaining a processing operation when normalizing a collation result between a feature vector extracted from a speech sound of a speaker to be recognized and a speaker model of the speaker.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Speech input part 2 ... Feature amount calculation part 3 ... Speaker model creation part 4 ... Speaker model storage part 5 ... Collation part 6 ... Normalization part

Claims (4)

[Claims] 1. A speaker recognition method for recognizing a speaker based on a speaker model created in advance from features extracted from the speaker's uttered sound. A speaker model having a tree-structured storage structure is created based on the obtained feature amount and stored in the storage means, and the feature amount extracted from the input utterance sound of the recognition target speaker and the stored speaker model are stored. And recognizing the speaker to be recognized. 2. A speaker recognition method for recognizing a speaker based on a speaker model created in advance from features extracted from the speaker's uttered sound. A speaker model having a tree-structured storage structure is created based on the obtained feature amount and stored in the storage unit. The feature amount extracted from the input utterance sound of the speaker to be recognized and the plurality of stored talks are stored. A speaker model used for normalizing a matching result between the input feature amount of the speaker to be recognized and the speaker model of the speaker to be recognized, The selected speaker model is used to normalize the matching result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker to recognize the recognition target speaker. Speaker recognition method. 3. A speaker recognition device for recognizing a speaker based on a speaker model created in advance from features extracted from the speaker's uttered sound. Storage means for creating and storing a speaker model having a tree-structured storage structure based on the obtained feature quantities, and feature quantities extracted from the input utterance sounds of the speaker to be recognized and stored in the storage means. A recognition unit for recognizing a speaker to be recognized by collating with a speaker model. 4. A speaker recognition device for recognizing a speaker based on a speaker model created in advance from feature amounts extracted from the speaker's uttered sound. Storage means for creating and storing a speaker model having a tree-structured storage structure based on the obtained feature quantities, and feature quantities extracted from the input utterance sounds of the speaker to be recognized and stored in the storage means. By comparing a plurality of speaker models, a speaker model to be used for normalizing a matching result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker is selected. By using the speaker model selected by the selecting unit, and normalizing the collation result between the input feature amount of the recognition target speaker and the speaker model of the recognition target speaker. And recognition means for recognizing the speaker to be recognized. Speaker recognition device, characterized in that was.