JPH05313697A - Speaker recognition system - Google Patents

Abstract

PURPOSE:To learn various phonemes by short vocalization by using a voice which has good phoneme balance. CONSTITUTION:A neural network is learnt by using the outline of a short-time spectrum obtained from a voice for learning. A voice with good phoneme balance is used as the voice for learning. For recognition, the outline of the same short-time spectrum is found from an optional speaking and its series is inputted to the network to obtain a network output. The output vectors of the obtained network output indicate speakers for the short-time inputs and are totally judged by sum, product, and majority decision making, etc., over the entire series to obtain one recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker recognition system (particularly speaker verification) using a neural network.

[0002]

2. Description of the Related Art Normally, in the case of speaker recognition, many speakers recognize only the voicing contents learned in advance.
Speaker recognition that does not limit the utterance content requires speaker utterance that is common to various phonemes, and thus requires a certain amount of utterance, and it is difficult to obtain a high recognition rate. Furthermore, since the amount of data during learning is very large, it took a long time for learning.

Therefore, the applicant has already proposed a speaker recognition method that does not limit the utterance content (Japanese Patent Application No. 2-243413 "Speaker Recognition Method", Japanese Patent Application No. 3-282843 "Speaker Recognition Method". ).

[0004]

However, the speaker recognition method proposed by the applicant of the present invention mainly identifies who the speaker is among a limited number of people (speaker identification). It was related to technology.

It is an object of the present invention to obtain a high recognition rate with a relatively short utterance in speaker recognition (especially speaker verification) without limiting the utterance content.

It is another object of the present invention to obtain a high recognition rate with a relatively short utterance and reduce learning in speaker recognition (especially speaker verification) without limiting the utterance content.

[0007]

A speaker recognition system according to claim 1 will be described. First, the neural network is learned using the outline of the short-time spectrum obtained from the learning voice. As the voice for learning, a phonologically balanced voice is used. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. The output vector of the obtained network suggests the speaker for each input for a short time.
One recognition result is obtained by making a comprehensive judgment such as a product or a majority vote.

A speaker recognition method according to claim 2 will be described. First, the neural network is learned using the outline of the short-time spectrum obtained from the learning voice. As the learning voice, a phonologically balanced voice is used. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained. Although the output vector of the obtained network suggests the speaker for each input for a short time, a threshold for output vector selection is provided,
One recognition result is obtained by selecting only a highly reliable output vector among these, and making a comprehensive judgment such as sum, product, and majority decision for all of these.

A speaker recognition method according to claim 3 will be described. First, the neural network is learned using the outline of the short-time spectrum obtained from the learning voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained.
The output vector of the obtained network suggests the speaker for each input for a short time, and by making a comprehensive judgment such as sum, product, majority vote, etc. in the whole sequence, one recognition result is obtained. To get

The speaker recognition method according to claim 4 will be described. First, the neural network is learned using the outline of the short-time spectrum obtained from the learning voice. At this time, the number of learning data is reduced by performing clustering for each speaker. At the time of recognition, an outline of the same short-time spectrum as described above is obtained from an arbitrary utterance, the sequence is input to the network, and a sequence of network output is obtained.
The output vector of the obtained network suggests a speaker for each input for a short time, but a threshold for output vector selection is provided, and only a reliable output vector is selected among them. For all these, sum,
One recognition result is obtained by making a comprehensive judgment such as a product or a majority vote.

The speaker recognition method according to claim 5 will be described. This is to set thresholds for speaker determination, which are used when making a determination regarding a speaker, one by one, and by using two thresholds, only the output corresponding to a certain speaker is activated. When the output value is large, the speaker is determined to be a registrant (see FIG. 1).

However, the description of the "neural network" in the present invention is as follows (1) to (4).

(1) The neural network has a structure similar to that of the hierarchical network shown in FIG.
It can be roughly classified into two types of mutual connection networks shown in (B). The present invention may be configured by using either of both networks, but the hierarchical network is more useful because a simple learning algorithm as described later has been established.

(2) Network Structure As shown in FIG. 3, the hierarchical network has a hierarchical structure including an input layer, an intermediate layer, and an output layer. Each layer is composed of one or more units. The coupling is only forward coupling such as input layer → middle layer → output layer, and there is no coupling in each layer.

(3) Structure of the unit The unit is a model of a brain neuron as shown in FIG. 4, and the structure is simple. Receive input from other units,
The sum is taken and converted according to a certain rule (conversion function), and the result is output. A variable weight, which represents the strength of the bond, is attached to each of the bonds with other units.

(4) Learning (Back Propagation) Learning a network is to bring an actual output close to a target value (desired output). Generally, the conversion function and weight of each unit shown in FIG. Is learned by changing.

As a learning algorithm, for example, Rumelhart, DE, McClelland, JL and the PDP Re
search Group, PARALLEL DISTRIBUTED PROCESSING, the
Backpropagation described in MIT Press, 1986. can be used.

[0018]

In the recognition method according to claim 1, the outline of the short-time spectrum used for learning is
It corresponds to various phonemes and transitions between phonemes, and by learning the neural network so as to obtain speaker information from these, any utterance can be dealt with.

By using a phonologically balanced voice, various phonemes can be learned with a short utterance.

By comprehensively judging the series of output vectors, even if there is an error in the judgment from one of the output vectors, the correct judgment can be made as a whole, and the recognition rate is improved.

Further, in the speaker recognition method according to the second aspect, by selecting a highly reliable output vector from among the output vectors, the comprehensive judgment becomes more reliable and the recognition rate is improved.

In the speaker recognition method of the third aspect, the outline of the short-time spectrum used for learning is
It corresponds to various phonemes and transitions between phonemes, and by learning the neural network so as to obtain speaker information from these, any utterance can be dealt with.

By comprehensively judging the series of output vectors, even if the judgment from one of the output vectors is incorrect, the correct judgment can be made as a whole, and the recognition rate is improved.

Since the representative vector of a plurality of data is used as the learning data by clustering, the number of learning data can be reduced while maintaining the learning effect. As a result, the learning time of the neural network can be greatly reduced. Furthermore, in the speaker recognition method according to claim 4,

By selecting a highly reliable output vector, the total judgment becomes more reliable,
The recognition rate is improved.

Further, in the speaker recognition method according to the fifth aspect, by setting the speaker determination thresholds one by one, the registered person and the non-registered person can be more accurately determined. The recognition rate is improved.

[0027]

【Example】

(First Example) For 5 registrants and 25 non-registrants, a short phonologically balanced short sentence for learning was Fourier-analyzed at a sampling frequency of 10 kHz, a frame length of 25.6 msec and a frame period of 12.8 msec. A series of 68ch (1/12 Oct.) power vectors is obtained in the band.

These vectors are used as inputs to the neural network (input layer 68 units, input pattern 1
Learn enough to activate the corresponding output unit only in the case of a registrant.

A sequence of power vectors is obtained in the same manner as for any utterance. This is input to the network learned by, and the sequence of output vectors {x ¹ , x ² , ..., X ⁿ } x ^t = (x ^t ₁ , ..., x ^t ₅ ) n: the number of frames is obtained.

The following three methods are used for the above vector series to determine whether the input is a registered person or a non-registered person.

(1) If the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds a preset threshold value for speaker determination, the registered person,
Otherwise non-registered person

(2) If the maximum value of Π _t x ^t _s (s = 1 to 5) exceeds the preset threshold for speaker determination, the registrant,
Otherwise non-registered person

(3) max {x ^t ₁ , ..., X ^t ₅ } = x ^t _s (s = 1
~ 5) is a registered person if the maximum value exceeds the preset threshold for speaker determination, otherwise a non-registered person

The following methods may be used instead of the above three methods (corresponding to claim 5).

(1) Only the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds the preset first speaker determination threshold value, and other values are preset. Registered person if it is less than the threshold for speaker determination of No. 2, otherwise non-registered person

(2) Only the maximum value of Π _t x ^t _s (s = 1 to 5) exceeds the preset first speaker determination threshold value, and other values are preset. Registered person if it is less than the threshold for speaker determination of No. 2, otherwise non-registered person

(3) max {x ^t ₁ , ..., X ^t ₅ } = x ^t _s (s = 1
The maximum value of the number of 5 to 5) exceeds the preset first speaker determination threshold value, and the other values have the preset second threshold value.
If it is below the speaker judgment threshold of

As an example of voluntary vocalization, a short sentence for learning, "He thought that the advance of science and technology and the courage of human beings have enabled a journey to a far universe." , "I'm home", "Hello" I was subjected to a speaker recognition experiments using the three-word of "Good morning", not used in the registrant five and learning using the learning non-subscribers
I was able to fully recognize 26 people.

(Second Embodiment) For 5 registrants and 25 non-registrants, sentences for learning were sampled at a sampling frequency of 10 kHz and a frame length of 25.6 msec.
Fourier analysis with a frame period of 12.8 msec, 100 to 5000 Hz
A series of 68ch (1/12 Oct.) power vectors is obtained in the band.

Hierarchical clustering is performed from these vectors to obtain about 200 representative vectors for each speaker.

These representative vectors are used as inputs to the neural network (input layer 68 units, input pattern is obtained by the number of speakers × the number of representative vectors after clustering), and the corresponding output unit is activated only in the case of a registrant. To learn enough.

For any utterance, a sequence of power vectors is obtained in the same manner as. This is input to the network learned by, and the sequence of output vectors {x ¹ , x ² , ..., X ⁿ } X ^t = (X ^t ₁ , ..., X ^t ₅ ) n: the number of frames is obtained.

The following three methods are used for the above vector series to determine whether the input is a registered person or a non-registered person.

(1) If the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds a preset threshold value for speaker determination, the registered person,
Otherwise non-registered person

(2) If the maximum value of Π _t x ^t _s (s = 1 to 5) exceeds the preset threshold value for speaker determination, the registrant,
Otherwise non-registered person

(3) max {x ^t ₁ , ..., X ^t ₅ } = x ^t _s (s = 1
~ 5) is the registered person if the maximum value exceeds the preset speaker judgment threshold value, otherwise the non-registered person

Further, the following method may be used instead of the above three methods (corresponding to claim 5).

(1) Only the maximum value of Σ _t x ^t _s (s = 1 to 5) exceeds the preset first speaker determination threshold value, and other values are preset. Registered person if it is less than the threshold for speaker determination of No. 2, otherwise non-registered person

(2) Only the maximum value of Π _t x ^t _s (s = 1 to 5) exceeds the preset first speaker determination threshold value, and other values are preset. Registered person if it is less than the threshold for speaker determination of No. 2, otherwise non-registered person

(3) max {x ^t ₁ , ..., X ^t ₅ } = x ^t _s (s = 1
The maximum value of the number of 5 to 5) exceeds the preset first speaker determination threshold value, and the other values have the preset second threshold value.
If it is below the speaker judgment threshold of

[0051]

As described above, according to the present invention, a high recognition rate can be obtained with a relatively short utterance in speaker recognition (especially speaker verification) without limiting the utterance content.

Further, according to the present invention, in speaker recognition (particularly speaker verification) in which the utterance content is not limited, a high recognition rate can be obtained with a relatively short utterance, and learning can be reduced.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing a threshold value for speaker determination and an output value of a network.

FIG. 2 is a schematic diagram showing a neural network.

FIG. 3 is a schematic diagram showing a hierarchical neural network.

FIG. 4 is a schematic diagram showing a structure of a unit.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI technical display location G10L 3/00 531 K 8842-5H

Claims (5)

[Claims] 1. A neural network is used to input a series of vectors representing the outline of a short-time spectrum, and the series of network outputs are integrated by summing, recognizing, and majority voting of recognition results by individual outputs. Is compared with a threshold for speaker determination to obtain a single recognition result, and a phoneme-balanced sentence is used as the network learning data. Speaker recognition method. 2. A neural network is used to input a series of vectors representing the outline of a short-time spectrum, and output vectors selected from the series of network outputs using output vector selection thresholds are output by individual outputs. A speaker recognition method characterized by obtaining one recognition result by summing the recognition results by sum, product, majority vote, etc., and comparing the result with a speaker judgment threshold value. A speaker recognition method characterized by using phonologically balanced sentences as data. 3. A neural network is used to input a series of vectors representing the outline of a short-time spectrum, and the series of network outputs are combined by summing, multiplying, majority voting, etc. of recognition results by individual outputs, and the result is obtained. Is compared with a threshold for speaker determination to obtain a single recognition result, and the speaker recognition is characterized by reducing the number of learning data of the network by clustering. method. 4. A neural network is used to input a series of vectors representing the outline of a short-time spectrum, and an output vector selected from the series of network outputs using an output vector selection threshold is output by individual outputs. A speaker recognition method characterized by obtaining one recognition result by summing the recognition results by sum, product, majority vote, etc., and comparing the result with a speaker judgment threshold value. A speaker recognition method characterized by reducing the number of data by clustering. 5. The speaker recognition method according to any one of claims 1 to 4, wherein the speaker determination thresholds are set one by one in magnitude.