JPH0415698A - Speaker collating system - Google Patents

Abstract

PURPOSE:To improve the collating rate by using a fresh neural network corresponding to the secular change of a voice when a neural network is restructured with input data for update learning. CONSTITUTION:The neural network is used to outputs a output value for decision making which indicates whether a current input speaker is registered or not. In this case, the neural network is restructured by learning based upon the learning input data on registered speakers. Then new input data on registered speakers are added to the existent learning input data to update the learning input data and the neural network is restructured by additional learning based upon the input data for update learning. Consequently, even if there is a period difference from the stage of the initial structuring of the neural network, the speaker matching system which does not deteriorate in collating rate is obtained.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a speaker verification system.

[Prior Art] The present applicant has proposed a speaker verification system that uses a neural network to output an output value for determining whether the current input speaker belongs to a registered speaker or an unregistered speaker. (Sound Picture Theory 2-6-4, PP, 53-54.19
89.3).

In this speaker verification system, a neural network is constructed by learning based on learning input data of registered speakers.

[Problems to be Solved by the Invention] However, in the prior art, the input data for learning is fixed only to data from a limited period.

For this reason, the voice input when performing verification may change over time compared to the voice used during learning, and as a result, the verification rate may deteriorate.

An object of the present invention is to provide a speaker verification system in which the verification rate does not deteriorate even if there is a time difference from the initial construction stage of the neural network.

[Means for Solving the Problem] The present invention according to claim 1 uses a neural network to obtain an output value for determining whether the current input speaker belongs to a registered speaker or an unregistered speaker. There is a speaker matching system that can output, and after building a neural network by learning based on the training input data of registered speakers, the new input data of registered speakers is accumulated on the existing training input data to generate the training input data. The neural network is reconstructed by updating the data and performing additional learning based on the input data for update learning.

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

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

■In the present invention, as shown in FIG. 3, existing learning input data of registered speakers or A 1 , A 2 in chronological order
, A 3 ...A n (A 1 is the oldest, An is the latest).

Then, if new input data A n+1 of the registered speaker is sursipulled, this A n+1 is accumulated in Al-An, and Al, A2, A3 = An, n of An+1.
+1 data is used as input data for update learning, and the neural network is reconstructed using this input data for update learning.

That is, the neural network is reconstructed using input data for update learning, and the matching rate can be improved by using a fresh neural network that corresponds to changes in speech over time.

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

■Since one or more of the elements of ■ to ■ according to claim 2 are used as input to the neural network,
Since the preprocessing for obtaining input is simple and the time required for this preprocessing is short, the speaker verification system can be easily processed in real time without using a complicated processing device.

[Example] Figure 1 is a block diagram showing the learning system of the neural network, Figure 2 is a block diagram showing the speaker verification system using the neural network, and Figure 3 is a schematic diagram showing input data for learning. .

(A) First, the learning system of the neural network will be explained (see Figure 1).

This system includes a voice input section 11, a preprocessing section 12, a learning data storage section 13, and a neural network 14.

The configurations of the preprocessing section 12, learning data storage section 13, and neural network 14 will be explained below.

(1) Preprocessing unit The preprocessing unit 12 performs simple preprocessing on input audio to create input data to the neural network 14.

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 plurality of (n) channel band pass filters 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 a certain 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) Learning data storage unit The learning data storage unit 13 stores learning input data.

At this time, the learning data storage unit 13 stores, as shown in FIG.
Update the learning input data of the registered speaker. In other words, existing human training data of registered speakers or AI, A2 in chronological order
, A3-An (AI is the latest 6, An is the latest). Then, if new input data A n+1 of the registered speaker is sampled, this data A n+1 is accumulated in Al to An, and the AI
, A2. A3... n+1 data of An and Arul are set as update learning input data.

(3) Neural Network The neural network 14 is constructed using the existing learning input data and the updated learning input data in the learning data storage section 13, and is used to determine whether the input speaker is an input speaker or a registered speaker. Output values can be output.

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

(2) The structural neural network 14 is, for example, a three-layer perceptron type, and the number of input units is 12n corresponding to the four blocks of the preprocessing section 12 and n channel, and the number of output units is the same as the number of registered speakers.

■The learning target value is:■For registered speakers, the output value of the corresponding output value is 1, and for other output values, it is 0.■For non-registered speakers, the output value of KanayamaKaninit is D.

(a) The voice of the registered speaker is subjected to preprocessing by the preprocessing unit 12 and is input to the neural network 14. The weights and conversion function of the neural network 14 are modified so as to approach the target values.

(The voice of the unregistered speaker is preprocessed by the preprocessing unit 12 and input to the neural network 14. The weight and conversion function of the neural network 14 are corrected so as to approach the target value.

(a) and (b) are the errors between the target value and the output value of the output unit, or are sufficiently small values (for example, I x 10-')
Repeat until it becomes.

(B) Next, a speaker verification system using a neural network will be explained (see FIG. 2).

This system includes the audio input section 11 and the preprocessing section 1 in (A) above.
2, a neural network 14, and a determination unit 15.

At this time, the determination unit 15 determines that the neural network software 14
The output pattern is transferred to neural network 1.
If the output value of any one of the four output units exceeds a certain threshold value and is close to 1°, the current input speaker is recognized as a registered speaker.

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

(1) Voice samples over a period of five months from the first month to the fifth month were used as input data for learning. Preprocessing was performed on 5 registered speakers and 25 non-registered speakers, and 64 dimensions (4 blocks x 16
channel) was obtained.

(2) A neural network was constructed by learning based on the input data for learning in (1) above.

(a) Those without additional learning: 25 yen per registered speaker using speech samples from the first month only.
This data was used as the input data for learning. A neural network was constructed by learning based on this learning input data.

(b) Additional learning (b-1) Speech samples from only the first month were used, and 25 pieces of data per registered speaker were used as input data for learning. A neural network was initially constructed by learning based on this learning input data.

(b-2) Create update learning input data by accumulating 5 newly sampled data every month after the first month into the existing learning input data, and create update learning input data based on this update learning input data. The neural network was rebuilt through additional learning.

(b-3) The additional learning described in (b-2) above was conducted a total of four times each month from the second to the fifth month.

(3) The evaluation data of registered speakers and non-registered speakers were input into each of the neural networks in (a) and (b) of (2) above, and judgments were made.

As a result, the matching rate of the neural network without additional learning was 90.2%, while the matching rate of the neural network with additional learning was 94.5%. That is, the present invention improved the matching rate by 4.3%.

In addition, the inputs to the neural network 14 created by preprocessing the input voice by the preprocessing unit 12 described above include: (1) temporal changes in the frequency characteristics of the voice, (2) average linear prediction coefficients of the voice, (2) Average PAR of audio (: OR coefficient, ■ Average frequency characteristics of audio and pitch frequency, ■ Average frequency characteristics of audio waveform with high frequency emphasis,
and ■ one or more of the average frequency characteristics of voice can be used.

The element of ■ above is the "temporal change in the average frequency characteristics of the voice within a certain time", the element of ■ above is the "temporal change of the average linear prediction coefficient within a certain time of the voice", and the element of the above The element of ■ is "temporal change in the average PARCOR coefficient within a certain time of audio", and the element of ■ above is "the average frequency characteristic and temporal change of pitch frequency within a certain time of audio", The above element (3) 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 (χn) of audio waveforms.

Therefore, assume that the following linear prediction is possible.

Linear predicted value χ, = -Σα1χt-1...(1) Linear prediction error εt=χ -χt...(2) Here, χt: sample value of the audio waveform at time t, (α+
) (i=1..=, p)= (quadratic) linear prediction coefficient Now, in the implementation of the present invention, the linear prediction error ε, 2
Find the linear prediction coefficient (α51) so that the root mean value is the minimum.

Specifically, (εt)2 is calculated, and its time average is (εt)
)2, a(εt)2/θα1=O, i=1.
2. ..., p, from the following equation (α,
) is required.

Furthermore, the PARCQR coefficient of (2) above is defined as follows.

That is, (kj (n=1..., p) is (9th order) PA
R(:OR coefficient (partial autocorrelation coefficient), PARC
The OR coefficient k nil is the forward residual ε♂ due to linear prediction
n and the backward residual ε,−1,. (The normalized correlation coefficient between b1 is defined by the following equation.

...(4) Here, ε, (fl: χ.

-Σ α 1Xt-i (α1): Forward prediction coefficient, t t-+n+1) Cb'= X t-(n+1+-
f, , i3...χt-J (β,): Backward prediction coefficient Further, the pitch frequency of the voice in the above (①) 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. .

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 speaker recognition.

On the other hand, if the average slope of the spectrum is not compensated for and the audio waveform is directly input to the neural network, the neural network will extract the feature of the average slope of the spectrum during learning. It takes time to extract the peaks and valleys of the spectrum necessary for speaker 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. or faster.

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

(2) The neural network 14 is reconstructed using input data for update learning, and the matching rate can be improved by using a fresh neural network 14 that responds to slight changes in voice.

■As an input to the neural network 14, one or more of the elements from ■ to ■ mentioned above, such as "temporal changes in the average frequency characteristics of audio within a certain period of time", are used, so it is difficult to obtain input. Since the preprocessing is simple and the time required for this preprocessing is short, the speaker verification system 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, it is possible to obtain a speaker verification system in which the verification rate does not deteriorate even if there is a time difference from the initial construction stage of the neural network.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a learning system of a neural network, FIG. 2 is a block diagram showing a speaker verification system using a neural network, and FIG. 3 is a schematic diagram showing input data for learning. 11... Voice input section, 12... Preprocessing section, 13... Learning data storage section, 14... Neural network, 15... Judgment section. Figure 1 Figure 2 Figure 3 (A) (B) Patent applicant Hiroshi Sekisui Chemical Co., Ltd. Representative 1) Kaoru Data

Claims (2)

[Claims] (1) A speaker verification system that uses a neural network to output an output value for determining whether the current input speaker belongs to a registered speaker or an unregistered speaker, and is used for learning registered speakers. After constructing a neural network through learning based on input data, the learning input data is updated by accumulating new input data of registered speakers on the existing learning input data, and the above is performed by additional learning based on the updated learning input data. A speaker matching system that reconstructs neural networks. (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 verification system according to claim 1.