JPH04295898A - Speaker collating system - Google Patents

Abstract

PURPOSE:To prevent only the category of a registerer from becoming large unnecessarily, and to improve the collation rate by generating newly a pattern of a non-registerer by a calculation and learning it, at the time of learning it additionally. CONSTITUTION:According to this speaker collating system, at the time of collating a speaker from an input voice by using a neural network 23, in the case of leaning additionally only a pattern of a registerer, an additional learning pattern of a non-registerer is not extracted from an actual voice sample but generated by a calculation, a candidate point of a non-registerer pattern is set by using a random number onto a multi-dimensional space by a result of analysis of a main component analysis executed to a registerer pattern, and whether the candidate point concerned is set as an actual non-registerer pattern or not is decided by whether a distance between the candidate point concerned and a category of each registerer is within a certain range or not.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker verification method suitable for verifying a speaker from input voice using an online terminal such as an electric lock or an IC card.

0002

[Prior Art] In general, patterns extracted from speech samples change over time. Therefore, the matching rate is high for speech samples immediately after learning, but there is a difference between the speech samples at the time of learning and the input speech samples. It is known that the matching rate deteriorates as the timing difference increases.

[0003] Therefore, the present applicant filed Japanese Patent Application No. 2-1208.
65 proposed a method for preventing deterioration of the matching rate in speaker matching using a neural network by adding patterns extracted from periodically new speech samples to the learning patterns and re-learning the trained neural network. There is.

[0004] However, when conducting additional learning in speaker verification using a neural network, registrants generally use the system many times over a long period of time, so they can prepare speech samples for additional learning, but non-registered persons Because it is difficult to prepare new audio samples,
In 865, the voice samples used for initial learning were repeatedly learned.

[0005] Also, the present applicant has filed Japanese Patent Application No. 2-24340.
7, in speaker verification using a neural network, a method has already been proposed in which learning patterns of non-registered persons are created by calculation instead of extracting them from actual speech samples.

[0006]

[Problems to be Solved by the Invention] However, in the prior art, as shown in FIG. 3, only the voice sample of the registrant is new, so the category occupied by the registrant in the pattern space becomes unnecessarily large, and the number of unregistered There is a risk that the voice samples of non-registrants may overlap with those of registrants (the shaded area in FIG. 3), and as a result, voice samples of non-registered persons may be erroneously matched with those of registrants. To prevent this, it is necessary to prepare voice samples of non-registered persons for additional learning. It was actually difficult to prepare. Furthermore, even if a voice sample could be prepared, it would not necessarily be possible to prevent unnecessary expansion of the registrant's category on the pattern space.

[0007] The present invention prevents only the category of registrants from becoming unnecessarily large and improves the matching rate by creating and learning patterns for non-registrants during additional learning. The purpose is to

[0008]

[Means for Solving the Problems] The present invention as set forth in claim 1 provides a method for additionally learning patterns of only registrants when verifying speakers from input speech using a neural network. This is a speaker matching method in which patterns are created by calculation rather than by extracting patterns from actual speech samples, and candidate points of non-registrant patterns are multidimensional based on the results of analysis such as principal component analysis performed on registrant patterns. Random numbers are set in the space, and it is determined whether the candidate point is an actual non-registrant pattern or not when the distance between the candidate point and each registrant category falls within a certain range. This is done depending on whether or not the person is present.

[0009] The present invention as set forth in claim 2 is the invention as set forth in claim 1, further comprising using a temporal change in frequency characteristics of audio as an input to the neural network.

[0010] The present invention as recited in claim 3 provides that when verifying speakers from input speech using a neural network,
When additionally learning only the patterns of registrants, the additional learning patterns of non-registrants are not extracted from actual speech samples, but are created by calculation, using a speaker matching method that uses the candidate points of the patterns of non-registrants as The candidate points are determined as grid points of a mesh in a multidimensional space based on the analysis results such as principal component analysis performed on the registrant pattern, and it is determined whether or not the candidate point is an actual non-registrant pattern. This is done based on whether the distance between the registrant and each registrant's category is within a certain range.

[0011] The present invention as set forth in claim 4 is the invention as set forth in claim 3, further comprising using a temporal change in frequency characteristics of audio as an input to the neural network.

[0012] In carrying out the present invention, the pattern used for additional learning for registrants and non-registrants may be only a new pattern, may be simply added to an old pattern, or may be a pattern used for additional learning of registrants and non-registrants. , old patterns may be deleted. The same applies to voice samples used for principal component analysis of registrants.

[0013]

[Function] According to the present invention as set forth in claims 1 and 3, the following effects (1) and (2) are achieved.

[0014] By adding a newly calculated pattern to the learning patterns of non-registrants, it is possible to prevent only the category of registrants from becoming unnecessarily large in the pattern space and improve the matching rate.

[0015] ■ There is no longer a need for voice samples for additional learning by non-registered persons.

According to the present invention as set forth in claims 2 and 4, the following effect (2) is achieved.

■Since "temporal changes in the frequency characteristics of audio" are used as input to the neural network, the preprocessing to obtain the input is simpler than the conventional complex feature extraction. The time required for processing is short. Therefore, speaker verification processing can be easily performed in real time without using a complicated processing device.

[0018]

[Embodiment] FIG. 1 is a schematic diagram showing a speaker verification system according to a first embodiment of the present invention, FIG. 2 is a schematic diagram showing a speaker verification system according to a second embodiment of the present invention, and FIG. 3 is a pattern diagram showing a speaker verification system according to a second embodiment of the present invention. It is a conceptual diagram of space.

(First Embodiment) (See FIG. 1) As shown in FIG. 1, the speaker verification system 10 includes a learning system and a verification system.

The learning system includes a voice input section 11, a preprocessing section 12, a principal component analysis section 13, a random number generation section 14, a distance calculation section 15, a pattern selection section 16, an inverse transformation section 17, and a learning pattern storage section 18. configured.

The verification system includes a voice input section 21, a preprocessing section 22, a neural network 23, and a determining section 24.

[0022] Hereinafter, each procedure of creating learning patterns for registrants, creating learning patterns for non-registrants, learning, and matching will be explained.

(A) Creation of learning pattern for registrant ■ The voice of the registrant is collected by the voice input section 11 and preprocessed by the preprocessing section 12 .

[0024] In this preprocessing, for example, the input audio "Tadaima" is temporally equally divided into four blocks and passed through a multi-channel band pass filter to obtain the frequency characteristics of each block, that is, each fixed time period. can be taken as a thing. At this time, the output of the bandpass filter is averaged by an averaging circuit for each block.

(2) The preprocessing result of (2) above is stored in the learning pattern storage section 18 as the registrant's learning pattern.

(B) Creation of learning patterns for non-registered users Creation of learning patterns for non-registered users depends on the arrangement of registrant patterns and non-registered user patterns on the feature space, but generally the dimensions of the feature space (input dimensions) ) is high, so it is performed on a reduced dimension feature space using principal component analysis or other transformations. In this example, the non-registered pattern was created by principal component analysis using the principal component analysis section 13.

(2) The principal component analysis section 13 subjects the registrant's learning pattern stored in the learning pattern storage section 18 to principal component analysis to obtain a transformation matrix.

[0028] In the random number generation unit 14, a pattern is set using random numbers on a multidimensional space defined by the principal component axes, and candidate points for learning patterns of non-registered persons are set.

■The distance calculation section 15 and the pattern selection section 16 select learning pattern data for non-registered users depending on whether the distance between the candidate point and each registrant's category is within a certain range. .

(2) In the inverse transformation unit 17, the product of the learning pattern data for the non-registered person selected in (2) above and the inverse matrix of the transformation matrix obtained by principal component analysis is taken as a learning pattern for the non-registered person.

② The creation result of ② above is stored in the learning pattern storage unit 18 as a learning pattern for a non-registered person.

(C) Learning The above (A) and (B) stored in the learning pattern storage unit 18
) The neural network 23 is additionally trained using the registrant pattern and non-registrant pattern as learning patterns.

(D) Verification The voices of registrants and non-registrants are collected by the voice input section 21, preprocessed by the preprocessing section 22, and the preprocessing results are input to the trained neural network 23. shall be. Thereby, the determination unit 24 performs speaker verification based on the output result of the neural network 23.

[0034] The preprocessing by the preprocessing unit 22 is, for example, similar to the preprocessing unit 12 in (A) above, to obtain the temporal change in the average frequency characteristic within a certain period of time of the audio. can do.

[0035] Experimental results using the speaker verification system 10 described above will be explained below.

■Using voice samples from 5 registered users (100 patterns) and 25 non-registered users (100 patterns),
Perform initial learning.

■ Once a month, registrant's voice samples (50
pattern) for additional learning.

■Additional learning patterns for non-registered users are created in the following manner.

(1) For the voice samples of 5 registrants,
Perform preprocessing to obtain a 64-dimensional feature vector.

(2) Perform principal component analysis on the feature vectors of the five people, and on the space spanned by the principal component axes up to the third axis,
Considering the distribution of registrant patterns, the non-registrant pattern (5
0 pattern).

■ Add the above 5 registrant patterns and the non-registrant pattern created by calculation to the learning pattern,
Retrain the trained neural network.

■ When the evaluation patterns of registrants and non-registrants were given as input to the trained network, the error rate was 20% lower than when the initial speech samples of non-registrants were repeatedly learned. Improvement was seen.

According to this first embodiment, by using the pattern of a registrant who is new in time and the pattern of a non-registered person created by calculation as the pattern for additional learning, the registrant's pattern is determined in the pattern space. It is possible to prevent only categories from becoming unnecessarily large and improve the matching rate. Furthermore, there is no need for voice samples for additional learning by non-registered persons.

(Second Embodiment) (See FIG. 2) The speaker verification system 30 differs from the speaker verification system 10 in that the random number generation section 14 is replaced by a grid point generation section 14A, as shown in FIG. Only in

That is, in the speaker verification system 30,
When creating a learning pattern for a non-registered person, when setting candidate points for the non-registered person pattern (step (■) in (B) above by the speaker matching system 10), it is necessary to This method sets grid points of a mesh in a multidimensional space as candidate points for learning patterns for non-registered users.

[0046] Note that the lattice points of a mesh are defined as a group of lattice planes arranged at a certain interval along each principal component axis, and a lattice plane group along each principal component axis direction is Refers to a lattice point formed by intersecting a group of lattice planes along the grid.

Therefore, in this second embodiment as well, by using the pattern of a chronologically new registrant and the pattern of a non-registered person created by calculation as the pattern for additional learning, it is possible to It is possible to prevent only the registrant category from becoming unnecessarily large and improve the matching rate. Also,
There is no longer a need for voice samples for additional learning by non-registered users.

[0048]

[Effects of the Invention] As described above, according to the present invention, when additional learning is performed, patterns of non-registrants are newly calculated and learned, thereby preventing only the category of registrants from becoming unnecessarily large. It is possible to prevent this and improve the matching rate.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram showing a speaker verification system according to a first embodiment of the present invention.

FIG. 2 is a schematic diagram showing a speaker verification system according to a second embodiment of the present invention.

FIG. 3 is a conceptual diagram of a pattern space.

[Explanation of symbols]

10, 30 Speaker verification system 11 Audio input section 12 Pre-processing section 13 Principal component analysis section 14 Random number generator 14A Lattice point generation section 15 Distance calculation section 16 Pattern selection section 17 Inverse conversion section 18 Learning pattern storage unit 21 Audio input section 22 Pre-processing section 23 Neural network 24 Judgment section

Claims (4)

[Claims] [Claim 1] When performing speaker verification from input speech using a neural network, if only the patterns of registered users are additionally learned, the additional learning patterns of non-registered users are not extracted from actual speech samples, but are computed. In the speaker matching method to be created, the candidate points of the non-registered person pattern are
Random numbers are set on a multidimensional space based on the analysis results such as principal component analysis performed on the registrant pattern, and a judgment is made as to whether or not the candidate point is an actual non-registrant pattern. A speaker verification method based on whether the distance between each registrant and the category is within a certain range. 2. The speaker verification method according to claim 1, wherein temporal changes in frequency characteristics of speech are used as input to the neural network. [Claim 3] When performing speaker verification from input speech using a neural network, if only the patterns of registered users are additionally learned, the additional learning patterns of non-registered users are not extracted from actual speech samples, but are computed. In the speaker matching method to be created, the candidate points of the non-registered person pattern are
Based on the analysis results such as principal component analysis performed on the registrant pattern, grid points of a mesh in a multidimensional space are determined, and it is determined whether or not the candidate point is an actual non-registrant pattern.
A speaker verification method based on whether the distance between the candidate point and each registrant's category is within a certain range. 4. The speaker verification method according to claim 3, wherein temporal changes in frequency characteristics of speech are used as input to the neural network.