JPH0415700A - Speaker recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to speaker recognition systems.

[Prior Art] The present applicant has proposed a speaker recognition system using a neural network. In this speaker recognition system, a certain threshold value θ is set for the output pattern of the neural network. For example, when the output value of one output unit of the neural network takes a value of 0 or more, If all of the output values are less than or equal to (1-θ), it is recognized that the current input speaker is the same speaker as the registered speaker corresponding to the output unit whose output value is 0 or more.

[Invention or problem to be solved] However, the output pattern of a neural network is not necessarily as described above, and there is a pattern in which there are two or more output units that exhibit an output value of θ or more, or ■ all outputs. Sometimes we see the appearance of a pattern in which the output value of the unit takes a value less than or equal to (l-θ). Such output patterns make it difficult or impossible to recognize the speaker.

An object of the present invention is to accurately perform speaker recognition even when the output pattern of a neural network constructed using all registered speakers as learning targets is a pattern that does not identify the input speaker.

[Means for Solving the Problems] The present invention according to claim 1 is a speaker recognition system using a neural network, which includes a neural network for all speakers constructed with all registered speakers as learning targets;
All registered speakers are divided into groups based on the fact that the output values of the output units that make up the neural network for all speakers have a certain similarity relationship with each other and cannot be judged by a threshold value.Learn the registered speakers for each group by dividing all registered speakers into groups. When the input speaker is recognized by the neural network for all speakers using the neural network for each group constructed as a target, and the output pattern of the neural network for all speakers is a pattern in which the input speaker cannot be identified. ,
Further, the present invention according to claim 2 is characterized in that a group including registered speakers corresponding to the output unit of the maximum output value in the neural network for all speakers is recognized by the group neural network. , As inputs to the neural network, ■temporal changes in the frequency characteristics of the voice, ■average linear prediction coefficients of the voice, ■average PARCOR coefficients of the voice, ■average frequency characteristics of the voice, and pitch frequency. , ■Average frequency characteristics of the audio waveform with high-frequency emphasis,
and (1) one or more of the average frequency characteristics of voice is used.

[Function] According to the present invention as set forth in claim 1, there is the following function and effect.

■By using a neural network for all speakers constructed using all registered speakers as learning targets and a neural network for each group constructed using registered speakers grouped by similar speakers as learning targets, It can improve the recognition rate. As a result, even if the output pattern of a neural network constructed using all registered speakers as learning targets is a pattern that cannot identify the input speaker, accurate speaker recognition can be performed.

According to the present invention as set forth in claim 2, there is the following effect (2).

■As input to the neural network, one or more of the elements of ■ to ■ according to claim 1 are used,
Since the preprocessing for obtaining input is simple and the time required for this preprocessing is short, the speaker recognition system can be easily processed in real time without using a complicated processing device.

[Example] Fig. 1 is a block diagram showing the learning system of the neural network for all speakers, Fig. 2 is a block diagram showing the learning system of the neural network for groups, and Fig. 3 is a block diagram showing the speaker recognition system. , FIG. 4 is a flowchart showing the speaker recognition system.

fA) First, the learning system of the neural network for all speakers will be explained (see Fig. 1).

This system includes a voice input section 11, a preprocessing section 12, and a neural network 13 for all speakers.

The configurations of the preprocessing unit 12 and the neural network for all speakers 13 will be described below.

(1) Pre-processing unit The pre-processing unit 12 performs simple pre-processing on input speech to create input data to the all-speaker neural network 13 and the group neural network 14 described below.

A specific example of the configuration of the preprocessing section 12 is as follows.

That is, the preprocessing section 12 may be a combination of a low-pass filter, a band-pass filter, and an averaging circuit.

■Cut the high-frequency noise components of the input audio signal using a low-pass filter. Then, this input audio is temporally equally divided into four blocks.

(2) Pass the audio waveform through a bandpass filter of multiple (n) channels to obtain frequency characteristics for each block, that is, for each fixed time period.

At this time, the output signal of the band pass filter is averaged for each block, that is, for each fixed period of time, in an averaging circuit.

Through the above pre-processing, the "temporal change in the average frequency characteristics of audio within a certain period of time" can be obtained.

(2) Neural network for all speakers The neural network for all speakers 13 determines whether the input voice is the input voice or all registered speakers.

A specific configuration of the neural network 13 for all speakers is as follows.

■Structure The neural network 13 for all speakers is, for example, a three-layer berceptron type, and the number of input units is 4 blocks of the preprocessing section 12, 4n units corresponding to n channels, and the number of output units is the same number as the number of registered speakers. .

■The learning target value is 1 for the output value of the corresponding output unit for the registered speaker, and 0 for the other output values.

(a) The voice of the registered speaker is subjected to preprocessing by the preprocessing unit 12 and is input to the neural network 13 for all speakers.

Modify the weights and conversion functions in [Neural Network for All Speakers] 3 so that they approach the target values.

(a) is repeated until the error between the target value and the output value of the output unit becomes a sufficiently small value (for example, I x 10-').

(B1 Next, the learning system of the group neural network will be explained (see Fig. 2).

This system includes a voice input section 11, a preprocessing section 12, a neural network for all speakers 13, a neural network for groups 14, a determination section 15, and a learning pattern storage section 16.
Consists of.

The configurations of the group neural network 14, determination section 15, and learning pattern storage section 16 will be described below. The preprocessing unit 12 is the same as the preprocessing unit 12 described above (A), and the neural network 13 for all speakers is the same as the preprocessing unit 12 described above (A).
) is used.

Each group neural network 14 is provided corresponding to each group that has been divided into groups in advance, and determines whether it is an input voice or a registered speaker in each group. Each group is defined as a group of speakers whose output values of the output units constituting the neural network 13 for all speakers have a certain similarity to each other and whose threshold value cannot be determined.
It is organized by grouping. Then, the input voices of the speakers forming each group are grouped as described above based on the determination result of the determination unit 15 and stored in the learning pattern storage unit 16, and the learning of the group neural network 14 is performed. provided for. Note that the learning pattern storage section 16 additionally includes a group storage section 16A.

Here, the above-mentioned "neural network 13 for all speakers"
A set of speakers whose output values of each output unit have a certain similarity relationship with each other and whose threshold value cannot be determined is defined as: ■ There are multiple units whose output values are equal to or higher than the threshold value of 0 in the output turn, or In a situation where threshold judgment cannot be made because all units take a value less than (1-θ), for example, the speaker corresponding to the output unit with the maximum output value and the speaker whose output unit has a certain difference from the maximum output value A speaker corresponding to an output unit whose output value is within a range (or a speaker corresponding to an output unit whose output value is within a range that is a certain ratio to the maximum output value)
A group of speakers consisting of

A specific example of the configuration of the group neural network 14 is as follows.

■The structural group neural network 14 is, for example, a three-layer perceptron type, and the number of input units is 2 in the preprocessing section.
There are 4 blocks, 4n units corresponding to n channels, and the number of output units is the same as the number of registered speakers constituting the group.

■The learning target value is the output value of the corresponding output unit for the registered speakers composing the group, and the other output values are 0.
shall be.

(a) The preprocessing unit 12 performs preprocessing on the voices of the registered speakers constituting the group, and inputs the preprocessed voices to the group neural network 14. The weights and conversion function of the group neural network 14 are corrected so as to approach the target values.

(a) is repeated until the error between the target value and the output value of the output unit becomes a sufficiently small value (for example, I x 10'').

(C) Next, the speaker recognition system according to the present invention will be explained (see FIGS. 3 and 4).

The speaker recognition system 10 includes a voice input section 11 and a preprocessing section 1.
2. Neural network for all speakers 13, neural network for group 14, and first determination unit 21,
It is composed of a group storage section 22, a network selection section 23, and a final determination section 24. At this time, the all-speaker neural network 13 was constructed in the above (A), and the group neural network 14 was constructed in the above (B).

The speaker recognition system 1o recognizes speakers using the following algorithms (1) to (3).

(1) The input speech is recognized by the neural network 13 for all speakers. The first determination unit 21 performs speaker recognition by thresholding the output patterns of the neural network 13 for all speakers, and identifies the input speaker.

(2) In (1) above, when the output pattern of the neural network 13 for all speakers is a pattern that does not specify the input speaker, the group storage unit 22 and network selection unit 23 are used to The group neural network 14 of the group including the registered speaker corresponding to the output unit with the maximum output value in the neural network 13 is selected.

(3) The input voice in (1) is recognized by the group neural network 14 selected in (2) above. The final determination unit 24 performs speaker recognition by thresholding the output pattern of the group neural network 14 and identifies the input speaker.

Hereinafter, specific implementation results of the speaker recognition system 10 will be explained.

(5) Using 5 people as registered speakers, preprocessing is performed on each speaker's voice sample to obtain a 64-dimensional (4 blocks x 16 channels) feature vector. Using this as input to the neural network, a neural network 13 for all speakers is constructed. As learning patterns, 5 registered speakers and 25 non-registered speakers were used.

■Groups for each registered speaker grouped based on the fact that the output values of the output units constituting the neural network 13 for all speakers have a certain similar relationship to each other and cannot be judged by a threshold value. A neural network 14 is constructed for this purpose.

Group 1 Speaker A and Speaker B Group 2 Speaker C and Speaker R etc. ② Using the neural network 13 for all speakers and the neural network 14 for group, speaker recognition is performed according to the above algorithm.

In the conventional case of only a neural network targeting all registered speakers, the output unit value is in the range of 1 to 1. ■The output value of output unit A is 0.1 for input speaker A.
7. The output value of output unit B is 0.19 ■For input speaker C, the output value of output unit C is 0.19, the output value of output unit D is 0.20, etc., making it difficult to recognize the speaker. Met.

In contrast, according to the speaker recognition system 10 of the present invention,
By using the group neural network 14 in addition to the all-speaker neural network 13, it has become possible to accurately recognize each of the registered speakers described above.

In addition, the inputs to the neural network created by preprocessing the input audio by the preprocessing unit 12 described above include: ■ Temporal changes in the frequency characteristics of the audio, ■ Average linear prediction coefficients of the audio, and ■ Audio. The average PARCOR coefficient of, ■The average frequency characteristic of the voice and the pitch frequency, ■The average frequency characteristic of the voice waveform with high-frequency emphasis,
and ■ one or more of the average frequency characteristics of voice can be used.

The element ■ above is the "temporal change in the average frequency characteristics of the audio within a certain time", the element ■ above is the "temporal change in the average linear prediction coefficient within the certain time of the audio", and the element The element (■) is "temporal change in the average PARCOR coefficient within a certain period of time", the element (■) above is "the average frequency characteristic and temporal change in pitch frequency within a certain period of time" (above). The element (2) can be used as a "temporal change in the average frequency characteristic within a certain period of time of a high-frequency emphasized audio waveform."

Incidentally, the linear prediction coefficient of (2) above is defined as follows.

That is, it is known that there is generally a high proximity correlation between sample values (χ) of audio waveforms.

Therefore, assume that the following linear prediction is possible.

Linear predicted value χt=-Σα1χ, -1...(1
) Linear prediction error ε, = χ, -χ, ... (2) Here, χt = sample value of the audio waveform at time t, (
α+) (1=L”..., p): (9th-order) linear prediction coefficient Now, in the implementation of the present invention, the linear prediction error ε, 2
Find the linear prediction coefficient (α) so that the root mean value is the minimum.

Specifically, (εt)2 is calculated and its time average is (ε,
)2, θ(εt)2/δα, :o, i=1.
2. ..., p, from the following equation (α1
) is required.

ΣCI Ivll-Jl = L J = ' + 2
+ “”+ p”” (3) Also, PAR (1
:The OR coefficient is defined as follows.

That is, [k, 1) (n=1..., p) (9th order)
When using PARCOR coefficient (partial autocorrelation coefficient), PAR
The COR coefficient k nil is the forward residual ε due to linear prediction
, <1) and the backward residual εt-fn+ll (b') is defined by the following equation.

ε (f)・εt−+n++1 ... (4) Here, ε 11: χt−Σ α・χ゛−”(α,)
: Forward prediction coefficient, (βj) : Backward prediction coefficient Further, the pitch frequency of the voice mentioned above (■) is the reciprocal of the repetition period (pitch period) of the vocal cord wave.

Furthermore, since we added pitch frequency, which is a basic parameter of the vocal cords that varies from person to person, as an input to the neural network, it is possible to improve the recognition rate of speakers, especially between adults/dwarfs and male/female. Ru.

Furthermore, the above-mentioned high frequency enhancement (2) is to compensate for the average slope of the spectrum of the audio waveform to prevent energy from concentrating in the low frequency range. However, the average slope of the spectrum of the speech waveform is common to all speakers and is unrelated to speaker recognition.

However, if the average slope of the spectrum is not compensated for and the audio waveform is directly input to the neural network, when the neural network learns, it will extract the feature of the average slope of the spectrum, and the speech will be distorted. It takes time to extract the peaks and valleys of the spectrum necessary for human recognition. On the other hand, when emphasizing the high frequencies of the input to a neural network, it is possible to compensate for the average slope of the spectrum, which is common to all speakers and has nothing to do with recognition, but which affects learning, which speeds up the learning process. becomes faster.

According to the above embodiment, there are the following effects (1) and (2).

■ By using a neural network 13 for all speakers constructed with all registered speakers as learning targets, and a neural network 14 for each group constructed with registered speakers grouped by similar speakers as learning targets. Speaker recognition rate can be improved.

As a result, even if the output pattern of a neural network constructed using all registered speakers as learning targets is a pattern in which the input speaker cannot be identified, accurate speaker recognition can be performed.

■As an input to neural network 13.14,
Since one or more of the above-mentioned elements (■ to ■), such as "temporal changes in the average frequency characteristics within a certain period of time", are used, the preprocessing to obtain input is simple; Since the time required for processing is short, the speaker recognition system 10 can be easily processed in real time without using a complicated processing device.

[Effects of the Invention] As described above, according to the present invention, even when the output pattern of a neural network constructed using all registered speakers as learning targets is a pattern in which the input speaker cannot be identified, speaker recognition can be performed accurately. It is possible to do this.

[Brief explanation of drawings]

Figure 1 is a block diagram showing the learning system of the neural network for all speakers, Figure 2 is a block diagram showing the learning system of the neural network for groups, Figure 3 is a block diagram showing the speaker recognition system, and Figure 4 is a block diagram showing the learning system of the neural network for all speakers. is a flowchart showing a speaker recognition system. DESCRIPTION OF SYMBOLS 10...Speaker recognition system, 11...Voice input unit, 12...Preprocessing unit, 13...Neural network for all speakers, 14...
・Group neural network, 21...1st
Judgment unit, 22...Group storage unit, 23...Network selection unit, 24...Final judgment unit. Patent applicant: Sekisui Chemical Co., Ltd. Representative Hiroshi 1) Kaoru

Claims (2)

[Claims] (1) In a speaker recognition system using a neural network, the output values of the all-speaker neural network built with all registered speakers as learning targets and the output units that make up the all-speakers neural network are mutually constant. Based on the fact that the threshold value cannot be determined due to the similarity relationship, all registered speakers are divided into groups, and a neural network for each group is constructed using the registered speakers of each group as learning targets. When recognition is performed using a neural network for all speakers, and when the output pattern of the neural network for all speakers is a pattern that cannot identify the input speaker, the output unit of the maximum output value in the neural network for all speakers is A speaker recognition system characterized in that a group including corresponding registered speakers is recognized by a group neural network. (2) As inputs to the neural network, [1] Temporal changes in the frequency characteristics of speech, [2] Average linear prediction coefficients of speech, [3] Average PARCOR coefficients of speech, [4] speech using one or more of the following: average frequency characteristics and pitch frequency; [5] average frequency characteristics of high-frequency emphasized audio waveform; and [6] average frequency characteristics of audio. The speaker recognition system according to claim 1.