JPH03157698A - Speaker recognizing system - Google Patents

Abstract

PURPOSE:To lessen the deterioration in a recognition rate with lapse of time and to allow real time processing with the recognizing system using a neutral network by making decision on a speaker and decision on the execution of additional learning in accordance with the thresholds for recognition of registered speakers and for additional learning. CONSTITUTION:The digital voice information from a voice processing section 12 is supplied to the neural network 13 and is stored into a memory 15. The recognition output value by this network 13 is compared with the threshold for deciding the speaker in a deciding section 14. Decision is made that the speaker is the registered speaker when the value is above the threshold. Further, the value is compared with the threshold for learning higher than this threshold. The learning is executed in accordance of the stored value of this time of the memory 15 via a network control section 16 when the value is below the threshold. The variable and coefft. of the network 13 are then changed and the network 13 is corrected to the network before the deterioration with lapse of time. The speaker recognizing system which lessens the deterioration in the recognition rate with lapse of time and allows the real time processing by using the neutral network is obtd. in this way.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speaker recognition system suitable for verifying a speaker from an input voice in an electronic lock or the like.

[Prior art] A conventional speaker recognition system is, for example, the Japanese Patent Publication No. 56-139
56, according to the following procedure.

■Extract features related to the speaker included in the input voice.

(2) Calculate the distance between the standard pattern extracted in advance in the same manner as (2) above and the feature amount extracted in (2) above.

(2) On the condition that the distance calculated in (2) above is smaller than a preset threshold, the current input speaker is determined to be the registered speaker of the standard pattern.

[Problem to be solved by the invention] However, in the conventional speaker recognition system described above, the following
, ■There are problems.

■As time passes from the time of standard pattern creation,
rate deteriorates. For example, after 3 months, the recognition rate is 1.
It deteriorates from 00.0% to 85.0%.

■Real-time processing is difficult. In other words, in order to secure a recognition rate higher than -room, it is necessary to use complex features;
In order to extract complex feature quantities, a complex processing device is required, and the processing time is also large.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speaker recognition system that exhibits extremely little deterioration in recognition rate over time and that can easily perform real-time processing.

[Means for Solving the Problems] The present invention according to claim 1 is a speaker recognition system using a neural network, which performs registered speaker recognition based on an output value of an output unit corresponding to a registered speaker. The current input speaker is determined to be a registered speaker, and the above output value is determined to be a registered speaker under the condition that the above output value is greater than the registered speaker recognition threshold. The current input voice data is used to perform additional learning of the neural network on the condition that the output value is greater than the registered speaker recognition threshold and smaller than the additional learning threshold. .

The present invention according to claim 2 provides, as inputs to the neural network, (1) temporal changes in the frequency characteristics of audio, (2) average linear prediction coefficients of audio, (2) average PARCOR coefficients of audio, and (2) the average PARCOR coefficient of audio. Average frequency characteristics and pitch frequency, ■Average frequency characteristics of high-frequency emphasized audio waveforms,
and (1) one or more of the average frequency characteristics of voice is used.

According to a third aspect of the present invention, the neural network is a hierarchical neural network.

[Effect] (1) Deterioration of recognition rate over time is extremely small. This has been confirmed by the experimental results described below, or it is presumed that the neural network can have a structure that is less susceptible to fluctuations due to differences in audio timing.

(2) In addition to the registered speaker recognition threshold, an additional learning threshold is provided for the output value of the output unit corresponding to the registered speaker, which constitutes the neural network. That is,
Even if the above output value exceeds the registered speaker recognition threshold and the input speaker can be determined to be a registered speaker, the output value exceeds the registered speaker recognition threshold. If it does not exceed a larger additional training threshold, the neural network performs additional training using the current input speech data.This prevents the recognition rate from deteriorating even if the speaker's characteristics change over time. As a result, it is possible to construct a speaker recognition system that is resistant to changes in speech over time.

(3) In principle, in a neural network, the calculation processing of the entire network is simple and quick.

(4) In principle, each unit constituting a neural network operates independently, and parallel arithmetic processing is possible. Therefore, calculation processing is quick.

(5) With the above (3) and (4), the speaker recognition system can be easily processed in real time without using a complicated processing device.

Further, according to the present invention as set forth in claim 2, the above (1) to (5)
), there is the following effect (6).

(6) Since one or more of each of the elements (■ to ■) described in claim 2 is used as input to the neural network, the preprocessing for obtaining the input is different from conventional complex feature extraction. Therefore, the process is simple, and the time required for this preprocessing is short.

Moreover, according to the present invention according to claim 3, the above (1) to (6)
In addition to the effects of ), there is also the effect of (7) below.

(7) Regarding hierarchical neural networks, a simple learning algorithm (pack propagation) as described later has been established, and a neural network that can achieve a high recognition rate can be easily formed.

[Example] Fig. 1 is a schematic diagram showing an example of a speaker recognition system to which the present invention is applied, Fig. 2 is a schematic diagram showing an example of a speech processing unit and a neural network, and Fig. 3 shows input speech. Schematic diagram, Figure 4 is a schematic diagram showing the output of the bandpass filter,
Figure 5 is a schematic diagram showing a neural network, Figure 6 is a schematic diagram showing a hierarchical neural network, and Figure 7 is a schematic diagram showing a hierarchical neural network.
The figure is a schematic diagram showing the structure of the unit.

Prior to describing specific embodiments of the present invention, the configuration of the neural network and the learning algorithm will be described.

(1) Due to its structure, neural networks are
The present invention can be roughly divided into two types: the hierarchical network shown in FIG. 5(A) and the interconnected network shown in FIG. 5(B). Hierarchical networks are more useful because simple learning algorithms have been established as described below.

(2) Network Structure A hierarchical network has a hierarchical structure consisting of an input layer, an intermediate layer, and an output layer, as shown in FIG.

Each layer is composed of one or more units. The connections are only forward connections such as input layer → middle layer → output layer, and there are no connections within each layer.

(3) Structure of the unit The unit is a model of a neuron in the brain and has a simple structure as shown in FIG. It receives input from other units, sums it up, transforms it using a certain rule (conversion function), and outputs the result. Each connection with another unit is given a variable weight that represents the strength of the connection.

(4) Learning (pack propagation) Learning of a network is to bring the actual output closer to the target value (desired output). Learn by making changes.

Further, as a learning algorithm, for example, Rumel
hart, D.E., McClelland, J.
L, and thePDP Re5earch Gr.
oup, PARALLEL DISTRIBUTED
PRO(:ESSING, the MIT Pres
s, 1986. Pack propagation as described in .

Hereinafter, specific examples of the present invention will be described.

As shown in FIG. 1, the speaker recognition system 10 includes a voice input section 11, a voice processing section 12, and a neural network 13.
, determination unit 14, memory unit 15, network control unit 16
, and a device control section 17.

(1) Input the registered voice to the voice input section 11.

At this time, the learning word is "Tadaima" and the input word is "Tadaima".

Also, there are 9 registered speakers and 27 impostors.

(2) The audio processing unit 12 performs simple preprocessing on the input audio in (1) above.

The preprocessing results are transferred to the neural network 13 for the current speaker recognition, and are also transferred to the memory unit 15 in preparation for the possibility of additional learning.

(3) The neural network 13 performs the following learning operation (2) and the following evaluation operation (2).

(2) The learning target value (target output value of each output unit constituting the output layer) is set to (1, 0) for the registered speaker and (0, 1) for the impostor.

The input voice "Tadaima" of the registered speaker is preprocessed by the voice processing unit 12, and the preprocessing result is input to the neural network 13. Then, the conversion function and weight of each unit of the neural network 13 are corrected so that the output value of the neural network 13 (the output value of each output unit constituting the output layer) approaches the target value.

This learning operation is repeated, for example, 30,000 times.

■Evaluation This time, the input voice of the speaker is preprocessed, and the preprocessed voice is input to the neural network 13 to obtain the output values (X, Y) of the neural network.

Then, the above output value (X
, Y) are transferred to the determination unit 14.

(4) The determination unit 14 determines the threshold value θ1.02.03 (
θ1>θ2).

Ol is a threshold for additional learning, θ2 is a threshold for registered speaker recognition, and θ3 is a threshold for impostor recognition.

The determination unit 14 performs the following determination operations (1) to (2) using the above threshold value.

(2) On the condition that [X>02 and Y<θ3], the determination unit 14 determines the current input speaker as a registered speaker, and outputs this registered speaker determination signal to the device control unit 17.

■[x>θ2 and Y>θ3 or [X<02 and Y>θ
3] or [X<02 and Y<θ3], the determination unit 14 determines the current input speaker to be an impostor, and outputs this impostor determination signal to the device control unit 17.

(2) Only when determining the registered speaker in (2) above, the determination unit 14 further performs the following processes (a) and (b).

(a) On the condition that [X<θ1], the determination unit 14 outputs an additional learning execution signal to the network control unit 16 in order to perform additional learning of the neural network 13 using the current input audio data. .

(b) [X>θl] When this is the case, the determination unit 14 does nothing.

(5) The device control section 17 controls the device using the registered speaker determination signal based on the determination result of the above-mentioned (2) by the determination section 14.

This device is, for example, an electronic lock, and performs unlocking control based on the registered speaker determination signal.

(6) The network control unit 16 determines to perform additional learning of the neural network 13 based on the additional learning execution signal based on the determination result of the above-mentioned (2) by the determining unit 14. At this time, the network control unit 16 controls the memory unit 1
5, extract the current input audio data, re-input this input audio data to the neural network 13, and calculate the output value of the neural network 13 for this input (X
, Y) or oral (3)■ Target value (1
, 0), the conversion function and weight of each unit of the neural network 13 are corrected. The network control unit 16 repeats this additional learning operation, for example, 30,000 times.

Hereinafter, as shown in FIG. 2, a specific example will be described in which a hierarchical neural network 13 is used and temporal changes in the average frequency characteristics of audio within a certain period of time are used as input to the neural network 13. do.

Note that the audio processing section 12 includes a low-pass filter 21, a band-pass filter 22, and an averaging circuit 23, as shown in FIG.
It is composed of the combination of

(2) The high-frequency components of the audio signal of the input audio are cut by the low-pass filter 21. Then, this input audio is temporally equally divided into four blocks as shown in FIG.

■The audio waveform is passed through a plurality (n) channel band pass filter 22 as shown in FIG. Obtain frequency characteristics.

At this time, the output signal of the bandpass filter 22 is averaged by the averaging circuit 23 for each block, that is, for a certain period of time.

Through the above preprocessing, the "temporal change in the average frequency characteristics of audio within a certain period of time" was obtained.

The output of the averaging circuit 23 is transferred directly to the neural network 13, or indirectly transferred to the neural network 13 via the memory section 15.

■The neural network 13 is composed of a three-layer hierarchical neural network. The input layer 3 consists of 4×n units corresponding to 4 blocks of preprocessing and n channels. The output layer 32 is composed of two units: a group of registered speakers and a group of impostors.

The target value of the output layer 32 is (1, Q
) for the imposter ((1, 1).

ExperimentAs mentioned above, as a result of comparing the recognition rate of the network that underwent additional learning by setting the threshold value θ1 for additional learning and the recognition rate of the network that did not perform additional learning, Table 1 was obtained. It is recognized that deterioration in recognition rate due to differences can be prevented.

Next, the operation of the above embodiment will be explained.

(1) Deterioration of recognition rate over time is extremely small. This has been confirmed by experimental results described below, and is presumed to be because the neural network 13 can have a structure that is less susceptible to fluctuations due to differences in audio timing.

(2) In addition to the registered speaker recognition threshold value θ2, an additional learning threshold value θl is provided for the output value of the output unit corresponding to the registered speaker that constitutes the neural network 13. In other words, even if the output value is greater than the registered speaker recognition threshold value θ2 and the input speaker can be determined to be the U-recorded speaker, the output value does not exceed the registered speaker recognition threshold value θ2. If the additional learning threshold θ1, which is larger than the additional learning threshold θ2, is not exceeded, the neural network 13 performs additional learning using the current input voice data.This allows the speaker's characteristics to change over time. However, the neural network 13 can be updated before the recognition rate deteriorates, and as a result,
It is possible to construct a speaker recognition system that is resistant to changes in speech over time.

(3) In principle, the neural network 13 has simple and quick arithmetic processing for the entire network.

(4) In principle, each unit constituting the neural network 13 operates independently, and parallel arithmetic processing is possible. Therefore, calculation processing is quick.

(5) According to (3) to (4) above, the speaker recognition system 1
0 can be easily processed in real time without using complicated processing equipment.

(6) As an input to the neural network 13, “
Because it uses "temporal changes in the frequency characteristics of the voice," the preprocessing to obtain input is complicated compared to conventional feature extraction.
It is simple and the time required for this preprocessing is short.

At this time, as an input to the above neural network,
Furthermore, since the "temporal change in the average frequency characteristic within a certain period of time" is used, the processing in the neural network 13 is simple, and the time required for this processing is shorter.

(7) Since the hierarchical neural network 13 is used, a simple learning algorithm that has already been established (
Pack propagation) can be used to achieve high recognition rates.

In the implementation of the present invention, as inputs to the neural network, ■temporal changes in the frequency characteristics of audio, ■average linear prediction coefficients of audio, ■average PARCOR coefficients of audio, ■average audio frequency characteristics and pitch frequency, ■Average frequency characteristics of high-frequency emphasized audio waveforms,
and ■ one or more of the average frequency characteristics of voice can be used.

Then, just as the element (■) above was further used as "temporal change in the average frequency characteristics within a certain period of time", the element (■) above is also used as "the average linear prediction coefficient within a certain period of time". ``time leopard change'', the element of ■ above is ``
``Temporal change in the average PARCOR coefficient within a certain time period of audio'', the above element (■) is ``the average frequency characteristic and temporal change in pitch frequency within a certain time period of audio'', and the above element (■) is: It can be used as [temporal change in the average frequency characteristics 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 -m has a high proximity correlation between sample values (χ) of audio waveforms.

Therefore, assume that the following linear prediction is possible.

Linear predicted value χ, = -Σα1χt-1...(1) Linear prediction error ε, =χ, -χL...(2) Here, χ: sample value of the audio waveform at time t, (α+
) (1=1..., p): (first-order) linear prediction coefficient Now, in the implementation of the present invention, 2 of the linear prediction error εt
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 (t
t)” and θ(t t)” / aa (=O
rz=1.2. By setting . . . , p, (a5) can be obtained from the following equation.

Further, the PARCOR coefficient of (2) above is defined as follows.

That is, [k,] (n=1...,p) is expressed as (first-order) P
When AR (:OR coefficient (partial autocorrelation coefficient)), PAR
COR coefficient k. , , is the forward residual ε due to linear prediction
<11 and backward residual ε′−(n*I)(b 1
The normalized correlation coefficient between is defined by the following formula.

...(4) Here, εt<1ゝ=χ is one Σ α direct χt−1゜(α arc): Forward prediction coefficient, εt−(n+II Lb'=χt−in+Il −F
, II J-χt-J.

(βj): Backward prediction coefficient Also, the pitch frequency of the voice mentioned in (2) above is the reciprocal of the repetition period (pitch period) of the vocal cord wave.

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

Furthermore, the above-mentioned high frequency enhancement (2) is to compensate for the average slope of the spectrum of the audio waveform to prevent concentration of energy 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 the speaker's recognition.

However, when inputting a speech waveform that has not been compensated for the average slope of the spectrum to a neural network as is, the neural network extracts the feature of the average slope of the spectrum during learning, and the speech becomes distorted. It takes time to extract the peaks and valleys of the spectrum necessary for human recognition. On the other hand, when high-frequency emphasis is applied to the input to a neural network, it is possible to compensate for the average slope of the spectrum that is common to all speakers and is unrelated to recognition, but that affects learning, which speeds up the learning process. becomes faster.

[Effects of the Invention] As described above, according to the present invention, it is possible to obtain a speaker recognition system that exhibits extremely little deterioration in recognition rate over time and that can easily perform real-time processing.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram showing an example of a speaker recognition system to which the present invention is applied, FIG. 2 is a schematic diagram showing an example of a speech processing unit and a neural network, FIG. 3 is a schematic diagram showing an input voice, and FIG. Figure 4 is a schematic diagram showing the output of the bandpass filter.
Figure 5 is a schematic diagram showing a neural network, Figure 6 is a schematic diagram showing a hierarchical neural network, and Figure 7 is a schematic diagram showing a hierarchical neural network.
The figure is a schematic diagram showing the structure of the unit. ○...Speaker recognition system, 1...Speech input section, 2...Speech processing section, 3...Neural network 4...Judgment section, 5...Memory section, 6...Network Control unit, 7... equipment control unit. twerk,

Claims (3)

[Claims] (1) A speaker recognition system using a neural network, which sets a registered speaker recognition threshold and an additional learning threshold for the output value of an output unit corresponding to a registered speaker, On the condition that the above output value is greater than the threshold for registered speaker recognition, the current input speaker is determined to be a registered speaker, and the above output value is greater than the threshold for registered speaker recognition, and additional information is added. A speaker recognition system that performs additional training of a neural network using the current input voice data, provided that it is smaller than a learning threshold. (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. (3) The speaker recognition system according to claim 1 or 2, wherein the neural network is a hierarchical neural network.