JPH02253298A - Voice pass filter - Google Patents

Abstract

PURPOSE:To provide the voice pass filter which can screen and remove a noise signal regardless of the kind of the noise signal superposed on a voice signal by providing a neural network for noise removal which screens and removes the acoustic parameter of the noise signal from the acoustic parameter of the inputted voice signal superposed with the noise signal. CONSTITUTION:The sampling value of the acoustic signal A consisting of the voice signal S superposed with the noise signal N is inputted via a delay section 20 to the input layer of the neural network 5 for noise removal. On the other hand, the environmental noise N is inputted mainly from a microphone 3 installed in the position apart at some extent from a speaker's mouth and is subjected to an A/D conversion in a sampling period of 12kHz by an A/D converter 4. The sampling value N(t) of the noise signal N is inputted via a delay section 21 to the input layer of the neural network 5 for noise removal. Then, the neural network 5 for noise removal separates the voice signal S and the noise signal N from the acoustic signal A, removes the noise signal N and outputs the voice signal S as an output sound.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to an audio pass filter that inputs an audio signal in which a noise signal is superimposed on an audio signal, removes the noise signal, and outputs only the audio signal.

<Prior art> Conventionally, as a method for separating an acoustic signal in which a noise signal is superimposed on a voice signal into a voice signal and a noise signal, spectrum analysis has been used.
There is a subtraction method. This spectral subtraction method is a method of suppressing noise signals and separating voice signals by subtracting the spectrum of silent sections (that is, only environmental noise) from the spectrum of voice signals on which noise signals are superimposed. There is also a method of estimating audio signal components by comparing simultaneous input audio signals from multiple microphones and separating the audio signals.

Furthermore, recently, attempts have been made to use neural networks to remove noise signals and separate speech signals from human acoustic signals (for example, Ogiyama and Itakura,
"Detection of speech intervals using neural networks" Proceedings of the Acoustical Society of Japan, 2-P-4, lθ, 1986).

Neural used to detect this speech interval.

The network is a three-layer berseptron type neural network, and the power of 19 frames and the linear prediction cepstrum up to the 12th order of three frames are input to the input layer. The intermediate layer is provided with five power-related units and O spectrum-related units. Further, there is one unit in the output layer, and the identification category is a voice section.

The training data uses the power of word sequence utterances of a specific speaker, white noise, printer noise, and linear predicted cepstrum. As a result of performing a speech segment identification test using a neural network trained using this training data, it was found that for acoustic signals such as word sequence utterances of a specific speaker of Ike and printer sounds similar to the above training data, You can improve your accuracy rate.

<Problems to be Solved by the Invention> However, in the above-mentioned speech segment detection using the neural network, one learning data is input for each training, so the trained neural - The network is structured to identify whether the input acoustic signal is a voice signal or a noise signal. Therefore, there is a problem in that the discrimination ability becomes considerably poor for an acoustic signal in which a noise signal having characteristics different from those of the learning data is superimposed on an audio signal.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide an audio pass filter that can select and remove a noise signal from an acoustic signal superimposed on an audio signal, regardless of the type of noise signal superimposed on the audio signal. It is about providing.

Means for Solving the Problems> In order to achieve the above object, the audio pass filter of the present invention includes a first acoustic analysis section that acoustically analyzes an audio signal on which a noise signal is superimposed to generate acoustic parameters; a second acoustic analysis section that acoustically analyzes a noise signal to generate acoustic parameters; and an acoustic parameter of an audio signal on which the noise signal output from the first acoustic analysis section is superimposed, and an acoustic parameter output from the second acoustic analysis section. The acoustic parameters of the noise signal are input, and through learning, the system automatically generates rules for selecting and removing the acoustic parameters of the noise signal from the acoustic parameters of the audio signal on which the noise signal is superimposed. The present invention is characterized in that it includes a noise removal neural network that selects and removes the acoustic parameters of the noise signal from the acoustic parameters of the superimposed audio signal according to the above rules.

Effects> The audio signal on which the noise signal is superimposed is input to the first acoustic analysis section and acoustically analyzed to generate acoustic parameters, while the noise signal is manually input to the second acoustic analysis section and acoustically analyzed. Acoustic parameters are generated. And the above first
Both acoustic parameters output from the acoustic analysis section and the second acoustic analysis section are manually input to a neural network for noise removal. Then, the noise removal neural network removes the acoustic parameters of the noise signal from the acoustic parameters of the audio signal on which the manually generated noise signal is superimposed, according to the rules it has generated by itself through learning, and outputs the result. Therefore, only the audio signal is selected and output from the audio signal on which the noise signal is superimposed.

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

FIG. 1 is a block diagram of the audio pass filter of the present invention. This audio pass filter has a noise removal neural network 5 having a structure as will be described in detail later, and outputs only the audio signal by manually inputting two audio signals.

In this embodiment, a signal mainly consisting of voice is referred to as a voice signal S, a signal mainly consisting of a noise signal is referred to as a noise signal N, and a voice signal S on which the noise signal N is superimposed is referred to as an acoustic signal A.

A speaker's voice on which environmental noise is superimposed is input from a close-talking microphone 1 placed near the speaker's mouth, and is A/D converted by an A/D converter 2 at a sampling period of 12 KHz.

Then, the sampling value A (tXt: sampling number) of the acoustic signal A obtained by superimposing the noise signal N on the audio signal S is input to the input layer of the noise removal neural network 5 via the delay unit 20, which will be described in detail later. be done. On the other hand, mainly environmental noise N is inputted from a microphone 3 installed at a position that can be seen from the speaker's mouth to some extent, and is A/D converted by an A/D converter 4 at a sampling period of 12 KHz. Then, the sampled value N (t) of the noise signal N is input to the input layer of the noise removal neural network 5 via the delay section 21 . Then, the noise removal neural network 5 separates the audio signal S and the noise signal N from the audio signal A, removes the noise signal N, and outputs the audio signal S as an output sound.

FIG. 2 is a schematic diagram of the structure of the neural network 5 for noise removal. This neural network consists of an input layer 6. Intermediate layer 7 and output layer 8
This is a three-layer perceptron type neural network with a three-layer structure consisting of: 2T units are arranged in the human power layer 6, N units are arranged in the intermediate layer 7, and T units are arranged in the output J18.

The 2T units of the input layer 6 are divided into T groups of two. Then, one unit 9, 10 . . . , 11 receives the sampling value A (t) of the acoustic signal A obtained by superimposing the noise signal N on the audio signal S from the A/D converter 2. Also, the other unit l of each group 2. + 3°..., 14,
The sampling value N(
Enter (). At that time, among the 2T groups, the t-th sampling value is input to the groups of units 9 and 12, and the (t
-1)th sampling value is input, and the (t-T +1)th sampling value is manually inputted to the group of unit 11.14 in the same manner. That is, in the input layer 6 of the noise removal neural network 5,
The acoustic signal A and the noise signal N from the (t-T+1)th sampling value to the tth sampling value are input.

In this case, A (t-T +1) to A (t) and N(
The method for inputting t-T+1) to N (t) is as follows. That is, the unit 9 includes the A/D converter 2.
The output signal A (t) from the unit 12 is directly inputted manually.
The output signal N (t) from the A/D converter 4 is directly input to the input signal. Further, the output signal A (t) from the A/D converter 2 is input to the unit 10 after being delayed by one sampling period by the delay element 22 of the delay section 20.
The unit I3 receives the output signal N (
t) is input after being delayed by the violet sampling period by the delay element 24 of the delay unit 21. Similarly, unit 11 receives output signal A from A/D converter 2.
(t) is input to the unit 14 with a time delay of (T-1) sampling period by the delay element 23, and the output signal N (t) from the A/D converter 4 is input to the unit 14 by the delay element 25 (T -1) The time may be delayed by the sampling period and then input.

Further, the t-th sampling value S (t) of the audio signal S is assigned to the unit 15 of the output layer 8, and the unit 1
6 indicates the (g)th sampling value 5() of the audio signal S.
t-t), and in the same manner, the unit 17 assigns the (t-T +l)th sampling value S of the audio signal S.
(t-741) is assigned. Each unit of the human power layer 6 is connected to all units of the intermediate layer 7, and each unit of the intermediate layer 7 is connected to all units of the output layer 8, respectively. However, units within each layer are not connected.

The learning of the noise removal neural network 5 is performed by the error backpropagation method as follows. That is, two types of learning data are prepared for this learning. The first learning data is an acoustic signal A in which a noise signal N° is superimposed on a voice signal S° of many speakers.

T sampling values A'(t-T+1)~A
Prepare (1). Also, as the second learning data,
T sampling values N' in the above noise signal N°
(t-T +l) to N'(t) are prepared. Also, as training data, T sampling values S' (
t-T +J) to S' (t) are prepared.

During learning, the first learning data is applied to each unit 9, 10 . . . of the input layer 6. . . , +1, while inputting the second learning data to each unit 12, 13, . . . of the input layer 6.
14, the teacher data is manually input to the output layer 8 as follows. That is, the unit 15 to which the t-th sampling value S (t) of the audio signal S is assigned the t-th sampling value T (t) = S'' of the teacher data.
(1) is input, and the (t-1)th sampling value T(t-1)=S'(t-1) of the teaching data is input to the unit 16 to which sampling value 5(t-1) is assigned. input the sampling value S (t-T +1
) is assigned to unit 17, the teacher data (t
−T +1)th sampling value T (t−T +l
) = S' (t-T +1). Then, the neural network for noise removal 5
is unit 15. of output layer 8. ...16, ..., 1
The network structure is determined by resetting the weights of the network so that the output value from 7 is the same as the teacher data.

In other words, this noise removal neural network 5
is learned by inputting a pattern (c) in which a pattern belonging to a certain category (a) is superimposed with another pattern (b) belonging to another category, and the above pattern (b), the pattern (b) is included. Select the pattern (C) from the pattern (C).
This is learning to find rules for selecting and eliminating (b). Therefore, the neural
The selection ability in the network is based on the above pattern (a).
It does not depend on the content of the bata-n (ro) superimposed on the. Therefore, in the noise removal neural network 5, the pattern (a) is the audio signal S and the pattern (b) is the audio signal S (the noise signal N superimposed on this). The neural network 5 can select and remove the noise signal N from the acoustic signal A, regardless of the type of the noise signal N.

The selection of the noise signal N from the input acoustic signal A by the noise removal neural network 5 trained as described above is performed as follows. That is, A/D converter 2
The sampling value A (t) of the acoustic signal A obtained by superimposing the environmental noise signal N on the audio signal S output from the delay unit 20
The time is delayed by (T-1) to 0 sampling periods, and the obtained sampling value A <t-T+1)
~A (t) is a neural network for noise removal 5
Each unit 9, 10 . . . . is input to II. On the other hand, the sampling value N (t) of the environmental noise signal N output from the A/D converter 4 is delayed by (T-1) to 0 sampling periods by the delay unit 21, and the obtained sampling value N , (t-T 41) ~N
(t) for each unit 12.13. . . . is input to 14. Then, the noise removal neural network 5 calculates the sampling value A (t-T +
l) - Sampling value N of environmental noise signal N from A(1)
Ct-T+l) to N(t) are selected and removed. The unit 15 of the output layer 8 outputs an output value corresponding to the sampling value S (t) of the audio signal S, and the unit 16 outputs an output value corresponding to the sampling value 5 (tl) of the audio signal S. and unit 1 in the same way.
7, the sampling value S (t−T +
1) is output.

Therefore, if the output signal from the unit 15 of the output layer 8 of the neural network 5 for noise removal is the output signal 0 (1) of the audio pass filter, the human layer 6 of the neural network 5 for noise removal The sampling values A (t-T +1) to A (t) of the signal A and the sampling values N (t-T 4
1)~N (t) is input, the unit of output layer 8! 5, the output signal 0 corresponding to the audio signal S (t)
(t) is output. Then, after the time period of ■sampling period has passed, the sampling value of the acoustic signal A input to the input layer 6 becomes A (t-T→2)=A(t+1),
Also, the sampling value of the environmental noise signal N is N (t-T
+2) to N(t+1).

The unit 15 of the output layer 8 outputs an output signal 0(t+1) corresponding to the audio signal S(t+I).

That is, the present audio pass filter uses a sampling value A (t
) and the sampling value N(t) of the noise signal N are input, the sampling value N(t) of the noise signal N is selected and removed from the sampling value AQ) of the audio signal A, and the sampling value N(t) of the noise signal N is input. It outputs an output signal 0(t) corresponding to the sampling value S(t).

In this manner, the audio pass filter of this embodiment obtains the sampling value A(t) of the audio signal A, in which the noise signal N is superimposed on the audio signal S, using the A/D converter 2, and performs sampling of the noise signal N. The value N (t) is determined by the A/D converter 4. Then, based on the sampling value A (t) of the acoustic signal A and the sampling value N (t) of the noise signal N, the noise removal neural network 5 extracts the sampling value A (t) of the acoustic signal A from the noise signal N (t). The sampling value N (t) is selected and removed, and an output signal 0(t) corresponding to the audio signal 5(t) is output.

Therefore, if the noise removal neural network 5 is properly trained using appropriate training data,
Regardless of the type of noise signal N superimposed on the audio signal S,
The audio signal S can be selected and output from the audio signal S on which the noise signal N is superimposed.

In the above embodiment, a three-layer perceptron type neural network is used as the noise removal neural network 5, but a four or more layer perceptron type neural network may be used.

In the above embodiment, the first learning data (acoustic signal A in which the noise signal N is superimposed on the speech signal S of many speakers) and the second learning data are input when the noise removal neural network 5 is trained. Learning data (the above superimposed noise signal N)
It is the same when sampling with. However, in reality, the noise signal input to the close-talking microphone 1 and the noise signal input to the microphone 2 have different phases, so the input timing of the second learning data is set to a predetermined value from the input timing of the first learning data. It may be delayed by a certain amount of time.

By doing so, it is expected that the selection ability of the noise removal neural network 5 will be further improved.In the above embodiment, the human input signal to the noise removal neural network 5 and the input signal from the noise removal neural network 5 are As an output signal, an A/D converter 2,
The sampled values A/D converted by 4 are used. However, the present invention is not limited thereto, and feature parameters such as an autocorrelation coefficient sequence and a spectral value sequence obtained by acoustic analysis of a speech signal or a noise signal on which a noise signal is superimposed may be used.

<Effects of the Invention> As is clear from the above, the audio pass filter of the present invention includes a first acoustic analysis section, a second acoustic analysis section, and a neural network for noise removal, so that a noise signal is superimposed. A first acoustic analysis section acoustically analyzes the generated audio signal to generate acoustic parameters, while a second acoustic analysis section acoustically analyzes the noise signal to generate acoustic parameters, and the noise removal neural network performs
The acoustic parameters of the noise signal output from the second acoustic analysis unit are determined from the acoustic parameters of the audio signal on which the noise signal output from the first acoustic analysis unit is superimposed, according to rules generated by the user through learning. Since the noise signal is removed and output, it is possible to select and remove only the noise signal from the audio signal on which the noise signal is superimposed, regardless of the type of noise signal superimposed on the audio signal.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of the audio pass filter of the present invention, and FIG. 2 is a schematic diagram of the structure of the noise removal neural network in FIG. 1. 1... Close-talking microphone, 2.4... A/D converter, 3.
...Microphone, 5...Neural network for noise removal, 6...
・Manpower layer, 7...middle layer, 8...output layer,
20.21...Delay section.

Claims (1)

[Claims] (1) a first acoustic analysis section that acoustically analyzes an audio signal on which a noise signal is superimposed to generate acoustic parameters; a second acoustic analysis section that acoustically analyzes the noise signal to generate acoustic parameters; The acoustic parameters of the audio signal on which the noise signal is superimposed output from the first acoustic analysis section and the acoustic parameters of the noise signal output from the second acoustic analysis section are input, and through learning, the audio on which the noise signal is superimposed is generated. It generates a rule for selecting and removing the acoustic parameters of the noise signal from the acoustic parameters of the signal, and selects the acoustic parameters of the noise signal from the acoustic parameters of the audio signal on which the input noise signal is superimposed according to the above rules. An audio pass filter characterized by comprising a neural network for removing noise.