JPH06282297A - Voice coding method - Google Patents

Abstract

PURPOSE:To provide a voice coding method by which the influence of a surrounding noise superimposed upon an input voice signal can be reduced. CONSTITUTION:A silence detecting part 3 to detect a silent interval of an input voice signal, an FFT part 4 to measure a spectrum pattern of the input signal, a noise spectrum pattern group 7 prepared beforhand, a neural network group 9 to learn extraction of a voice signal with every noise pattern in this group, and a coding part 12, are provided, and an input pattern with every detection of the silence timing and a prepared noise pattern are compared with each other, and the closest noise pattern is selected, and an input signal converted into a frequency area is inputted to a neural network learnt beforehand by this selected noise pattern, and is coded after the voice signal is extracted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice encoding system for digitizing a voice signal for transmission or storage and converting the transmitted or stored digital signal into an analog signal, such as a telephone, a mobile telephone, an automobile. The present invention relates to a voice encoding method applicable to telephone devices such as telephones, voice files, voice memories, and the like.

[0002]

2. Description of the Related Art Conventionally, this kind of speech coding system has a function of measuring the amplitude of an input speech signal and controlling the amplitude of the inputted speech in order to reduce deterioration of the speech signal due to ambient noise. It has a function to do so, detects a portion where the amplitude of the input audio signal is lower than a certain threshold value (hereinafter referred to as "silent section"), and attenuates the signal level of this silent section. This method is intended to reduce the influence of ambient noise by controlling so that the situation where only ambient noise is encoded and emphasized when there is no voice input at all will occur. there were. FIG. 5 is a diagram showing the configuration of a conventional speech coding method with noise countermeasures. In FIG. 5, reference numeral 51 is an audio signal input, which is connected to the A / D conversion unit 52. A / D converter 5
The output of 2 is connected to the silence detector 53, the changeover switch 67, and the encoder 62. The changeover switch 67 is connected to the attenuator 68. The output of the encoding unit 62 is
It is connected to the decoding unit 64 via the transmission line 63. The output of the decoding unit 64 is connected to the D / A conversion unit 65 and
The output of the A converter 65 is connected to the audio signal output terminal 66. With the above configuration, an analog audio input signal input from the audio signal input terminal 51 is converted into a digital signal by the A / D converter 52, and then the silence detector 53.
Thus, the silent section is detected. Then, in the silent section, the changeover switch 67 is turned on, and the attenuator 6
Since the audio signal is grounded via 8, the level of the input signal is lowered. With the above-mentioned mechanism, it has been prevented that only the noise signal is encoded and only noise is conspicuous when the input voice signal is silent.

[0003]

However, the above-mentioned conventional speech coding system is a noise reduction measure only when there is no input voice signal and there is no input voice signal. It was not possible to reduce the effect of noise on the. The present invention has been made to solve the above problems, and an object of the present invention is to provide a speech coding method capable of reducing the influence of ambient noise superimposed on an input speech signal.

[0004]

In order to solve the above problems, a speech coding system according to the first invention of the present application is a silent section detecting means for detecting a silent section of an input speech signal, and the input speech signal. Input spectrum pattern measuring means for measuring the spectrum pattern of the above, a noise spectrum pattern group prepared in advance, a neural network for learning voice signal extraction for each noise spectrum pattern in the noise spectrum pattern group, and an encoding means. And, for each timing detection of the silent section, the spectrum pattern of the input audio signal at that time is compared with individual noise spectrum patterns in the noise spectrum pattern group, and the closest noise spectrum pattern is selected. Two previously learned with the selected noise spectrum pattern Enter the input audio signal transformed into the frequency domain to over neural network, configured to perform encoding after the extraction of the audio signals only. Also,
A voice encoding system according to a second invention of the present application comprises a silent section detecting means for detecting a silent section of an input speech signal, an input spectrum pattern measuring means for measuring a spectrum pattern of the input speech signal, and a time measuring means. A time zone spectrum pattern accumulating means for accumulating a time zone spectrum pattern which is a spectrum pattern during a silent period controlled by the time measuring means, a comparison voice signal recorded in a noiseless state, and a learning function A neural network group and an encoding unit are provided, and learning is performed in advance to extract a voice signal by the sum of the time period spectrum pattern and the comparison voice signal, and voice is adaptively applied to noise that changes depending on the time period. It is configured to perform encoding after performing only signal extraction.

[0005]

According to the speech coding method of the first invention of the present application having the above-mentioned configuration, when the amplitude of the input speech signal is equal to or less than a certain threshold value, that is, in the silent section, the spectrum analysis of the ambient noise is performed and input. The spectrum pattern is measured, and the noise spectrum closest to the measured input spectrum pattern is selected from the noise spectrum pattern group prepared in advance. On the other hand, a neural network that has received a signal in which an assumed assumed noise spectrum pattern and a voice signal are mixed and learned by using only the voice signal as a teacher signal has a capability of extracting the voice signal in the mixed signal. Therefore, by preparing a neural network group that has learned the voice extraction with the noise spectrum pattern group consisting of a plurality of noise spectrum patterns prepared in advance as the above-mentioned assumed noise spectrum pattern, Effects of ambient noise by selecting one of the prepared noise spectrum patterns, selecting a neural network that has been pre-learned with the noise spectrum pattern, inputting an input signal to the neural network, and then encoding It is possible to perform speech coding excluding. Further, according to the speech coding method according to the second invention of the present application having the above-mentioned configuration, the input spectrum pattern of the ambient noise is measured and accumulated at constant time intervals, that is, at each time interval, and the accumulated time interval is measured. Since the neural network that performs learning using the noise spectrum pattern and the comparative speech signal is provided, the speech signal can be extracted by adapting to the characteristics of the ambient noise for each time zone of the place where the speech encoding device is installed. it can. By encoding the extracted signal, it is possible to more effectively remove the influence of ambient noise and perform voice encoding.

[0006]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows the configuration of a speech coding system which is a first embodiment of the present invention. In FIG. 1, reference numeral 1 is an audio signal input terminal, which is connected to the A / D converter 2. The A / D conversion unit 2 is a silence detection unit 3 which is a silent section detection unit and an F which is an input spectrum pattern measurement unit.
It is connected to an FT (Fast Fourier Transform) unit 4. The output of the silence detector 3 is connected to the neural network first selection switch 8. Also,
The output of the FFT unit 4 is connected to the changeover switch 5 and the neural network first selection switch 8 described above.

The changeover switch 5 is a switch which is controlled by the silence detecting section 3 so as to be turned on when there is no sound, and its output is connected to the selecting section 6. A noise spectrum pattern group 7 is connected to the selection unit 6. Thereby, the selection unit 6 selects the closest one of the spectrum patterns prepared in advance in the noise spectrum pattern group 7 to the spectrum pattern given from the FFT unit 4 when there is no sound, and based on the result,
The neural network first selection switch 8 and the neural network second selection switch 10 are controlled.

Reference numeral 9 is a neural network group which has been preliminarily learned to extract a voice signal from a signal in which a noise spectrum pattern and a voice signal which are prepared in advance are mixed, and are provided with a plurality of neural network units 91 to 9n. There is. Each neural network unit has an FFT unit 4
The input voice signal converted into the frequency domain by is switched by the neural network first selection switch 8 and input.

The output of the neural network group 9 is switched by the neural network second selection switch 10 controlled by the selection unit 6 to change the IFFT.
It is output to the (inverse fast Fourier transform) unit 11. IFFT
The unit 11 transforms the input signal into a time domain signal, and then outputs it to the encoding unit 12, which is an encoding means. The output coded by the coding unit 12 is input to the decoding unit 14 via the transmission line 13, is decoded, and is then converted again into an analog signal by the D / A conversion unit 15 to output a voice signal. It is output from the terminal 16.

Next, the operation of the first embodiment will be described. The analog audio signal input from the audio signal input terminal 1 is converted into a digital signal in the A / D converter 2. The digitized signal is input to the silence detector 3 and the timing of the silent section is detected. This output is sent to the changeover switch 5 described above, and a control signal for turning on the changeover switch 5 is sent from the silence detecting section 3 to the changeover switch 5 while the silent section is detected. While the changeover switch 5 is ON, the input signal converted into the frequency domain by the FFT unit 4 is input to the selection unit 6 through the changeover switch 5.

The selecting section 6 compares the spectrum pattern prepared in advance in the connected noise spectrum pattern group 7 with the input signal, and selects among the spectrum patterns prepared in advance in the noise spectrum pattern group 7. Select the closest one. As a result, it is possible to detect what kind of noise the input signal in the silent section, that is, the spectrum of the ambient noise.

An example of spectrum patterns prepared in advance in the noise spectrum pattern group 7 is shown in FIG.
In FIG. 2, FIG. 2A is an example of office noise such as an office, FIG. 2B is an example of wind noise, and FIG. 2C is an example of noise inside a vehicle. Shows. In addition to these, it is possible to prepare various noise patterns.

FIG. 3 shows a more detailed structure of a portion related to the neural network unit 9 described above. As shown in the figure, in a certain neural network unit 9m, a noise signal 22 prepared in the noise spectrum pattern group 7 and having the same noise spectrum as that selected by the selecting unit 6,
The signal obtained by adding the input audio signal 21 and the adder 23 is converted by the FFT unit 4b into a frequency domain signal analyzed into a frequency band of 256 channels, for example, and is normalized by the normalizing unit 24b and input. The output of the network unit 9m is input to the comparator 25. on the other hand,
Only the voice signal is converted into a frequency domain signal by the FFT unit 4a, and the signal normalized by the normalization unit 24a is input to the comparator 25 as a teacher signal.

The neural network unit 9m is
256 according to the number of FFT analysis channels
Each unit has an input layer of units, an intermediate layer of, for example, 200 units, and a network having a three-layer structure of an output layer of, for example, 256 units, which is adapted to the number of input units, and learning of voice signal extraction is performed.

Next, the operation at the time of learning with the above configuration will be described. The input voice signal 21 in FIG. 3 is a voice signal without noise, and voice signals of each age group of men and women recorded in advance in a noiseless environment are used. As the noise signal 22 in FIG. 3, assumed noise such as office noise and vehicle interior noise as shown in FIG. 2 is selected.

The voice signal 21 and the noise signal 22 are added by the adder 23 to generate a signal in which voice and noise are mixed. This adder output signal becomes a signal in the frequency domain normalized by the spectrum level for each frequency analysis frame in the FFT unit 4b and the normalization unit 24b.

On the other hand, after the signal of only the audio signal 21 is normalized in the FFT unit 4a and the normalizing unit 24a by the spectrum level of each frame, it is input to one of the comparators 25 as a teacher signal. Then, the output of the output layer of the neural network unit 9m is connected to the other side of the comparator 25. Then, the frequency domain signal of the above-mentioned signal in which the noise signal and the voice signal are mixed is connected to the neural network input layer of the neural network unit 9m so that only the voice signal as the teacher signal is output from the output layer. Is learned and the weighting factors wij and wjk of each unit are determined. In this case, the configuration of the neural network 9m is the FFT units 4a and 4b.
In accordance with the frequency resolution of, for example, a network configuration having an input layer of 256 units and an output layer of 256 units as described above.

The neural networks 91 to 9n shown in FIG.
As the above, the one learned as described above is used. When the noise spectrum pattern in the silent state is selected from the spectrum patterns of the noise spectrum pattern group 7 by the selection unit 6, the switching control of the neural network first selection switch 8 and the neural network second selection switch 10 is performed. Were, selection section 6
The neural network unit that performs the above learning (learning to extract a voice signal from a signal in which a voice signal and a noise signal are mixed) in advance according to the noise spectrum pattern selected by the plurality of neural network units 91 to 9n.
To be selected from. Therefore, after the switches 8 and 9 are switch-controlled in this way, the audio signal component extracted from the mixed signal is input to the IFFT unit 11.

This signal component is transmitted to the IFFT section 11
Is inverse-Fourier-transformed into a time domain signal, encoded by the encoding unit 12, sent to the decoding unit 14 via the transmission path 13, and decoded, and then the D / A conversion unit 1
It is converted into an analog signal again by 5 and output from the audio signal output terminal 16.

Thus, according to the first embodiment described above,
When the silence detector 3 detects a silent section of the input voice, the noise characteristic is determined to be one of the already measured patterns by comparing the spectrum pattern of the period with a noise spectrum pattern assumed in advance, and the noise characteristic is determined. A neural network that has already been learned in accordance with the above and has the ability to extract speech from a speech and noise mixed signal is specified, and the input speech signal is passed through the neural network, so that noise-induced degradation of the speech signal at the time of coding can be prevented. It is possible to suppress and suppress the influence of noise.

Next, a second embodiment of the present invention will be described with reference to FIG. In the second embodiment of FIG. 4, a noise spectrum pattern group corresponding to the noise spectrum pattern group 7 of FIG. 1 is added to the configuration similar to that of the first embodiment of FIG. In addition to (not shown), a time zone spectrum pattern group 46 which is a time zone spectrum pattern accumulating means is added, a time zone neural network group 47 is added, and further, a comparison voice signal 41 recorded in a noiseless state, F that performs FFT conversion of the signal
The FT unit 34a is provided and the time zone neural network is provided with a self-learning function.

Next, the operation of the above second embodiment will be described. The basic operation of the second embodiment is similar to that of the first embodiment, and the following operation is added.

First, by providing the timer 48, the average spectrum pattern for each time zone of the signal input at the timing of the silent section is accumulated in the time zone spectrum pattern group 46. Further, by the neural network prepared for each time zone of the time zone neural network group 47, the noise signal for each time zone accumulated in the above time zone spectrum pattern group 46 and the noise signal prepared in advance are recorded. FFT the compared audio signal 41
Learning is performed using the signal converted into the time domain signal by the unit 34a.

This learning is started after the noise spectrum for each time zone is accumulated to some extent, and the learning is ended by the voice signal extracted from the noise compared by the comparator 45 and the voice signal 41 for comparison. The time when the difference between the comparison results with and falls below a certain value.

After the above learning is completed, the selection range of the noise spectrum is expanded from only the prepared noise spectrum pattern as in the first embodiment to the prepared noise spectrum pattern and the time zone noise pattern. Alternatively, the deterioration of the voice signal due to noise is reduced by switching the selection method from the selection of the neural network by selecting the noise spectrum pattern to the selection of the neural network by the time zone.

Thus, according to the second embodiment described above,
Since the measurement of the noise spectrum pattern is performed at the place where the voice encoder is installed, the adaptive noise is used especially for the deterioration of the voice signal due to the noise of the voice encoder used at a fixed position. The influence of noise can be efficiently reduced by using the learned neural network. In general, the ratio of the silent section in the input speech reaches about 30%, so that measuring every silent period is equivalent to measuring the ambient noise almost every moment, and the speech coding apparatus has It becomes possible to fully adapt to changes in the installed acoustic environment.

Regarding the encoding unit 12 in the above embodiment, an ADPCM (Adaptive Differential Pulse C
ode Modulation), CELP (Code Excited Linear Pr)
ediction), LPC (Linear Prediction Coding) vocoder, and other various encoders can be used. Also, FIG.
The transmission line 13 in FIG. 2 can be of various types, such as a wired transmission line, a wireless transmission line, or a system in which it is stored in the memory and then reproduced.

The present invention is not limited to the above embodiment. The above-mentioned embodiment is an exemplification, has substantially the same configuration as the technical idea described in the scope of the claims of the present invention, and has any similar effect to the present invention. It is included in the technical scope of the invention. For example, in the above embodiment, an example in which noise is reduced in the encoding unit has been described, but the concept of noise reduction described above can be implemented not only in the encoding unit but also in the decoding unit.

[0029]

As described above, according to the speech coding method of the first invention of the present application having the above-mentioned configuration, when the amplitude of the input speech signal is equal to or less than a certain threshold, that is, in the silent section, the surrounding The noise spectrum analysis is performed to measure the input spectrum pattern, and the noise spectrum closest to the measured input spectrum pattern is selected from the noise spectrum pattern group prepared in advance.
On the other hand, a neural network that has received a signal in which an assumed assumed noise spectrum pattern and a voice signal are mixed and learned by using only the voice signal as a teacher signal has a capability of extracting the voice signal in the mixed signal. Therefore, by preparing a neural network group that has learned the voice extraction with the noise spectrum pattern group consisting of a plurality of noise spectrum patterns prepared in advance as the above-mentioned assumed noise spectrum pattern, Effects of ambient noise by selecting one of the prepared noise spectrum patterns, selecting a neural network that has been pre-learned with the noise spectrum pattern, inputting an input signal to the neural network, and then encoding It has an advantage that speech coding can be performed except for. Further, according to the speech coding method according to the second invention of the present application having the above-mentioned configuration, the input spectrum pattern of the ambient noise is measured and accumulated at constant time intervals, that is, at each time interval, and the accumulated time interval is measured. Since the neural network that performs learning using the noise spectrum pattern and the comparative speech signal is provided, the speech signal can be extracted by adapting to the characteristics of the ambient noise for each time zone of the place where the speech encoding device is installed. it can. By encoding the extracted signal, it is possible to more effectively remove the influence of ambient noise and perform voice encoding. Therefore, when the speech coding apparatus is fixedly used, there is also an advantage that deterioration of the speech signal due to ambient noise can be more effectively reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing an overall configuration of a speech coding system which is a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of noise patterns prepared in the noise spectrum pattern group shown in FIG.

3 is a block diagram showing a more detailed configuration of a portion related to the neural network unit shown in FIG.

FIG. 4 is a block diagram showing an overall configuration of a speech coding system which is a second embodiment of the present invention.

[Fig. 5] Fig. 5 is a diagram showing the configuration of a conventional speech encoding system with noise countermeasures.

[Explanation of symbols]

 1 Audio signal input terminal 2 A / D conversion unit 3 Silence detection unit 4, 4a, 4b FFT unit 5 Changeover switch 6 Selection unit 7 Noise spectrum pattern group 8 Neural network first selection switch 9 Neural network group 91 to 9n Neural network unit 10 Neural network second selection switch 11 IFFT unit 12 Encoding unit 13 Transmission path 14 Decoding unit 15 D / A conversion unit 16 Audio signal output terminal 21 Audio signal 22 Noise signal 23 Adder 24a, 24b Normalization unit 25 Comparator 32 A / D conversion unit 33 Silence detection unit 34a, 34b FFT unit 36 Selection unit 38a, 38b Changeover switch 39 Neural network unit 41 Voice signal for comparison 45 Comparator 46 Time zone spectrum pattern group 47 Time zone neural network group 48 Ma 51 audio signal input terminal 52 A / D converter 53 the silence detector 62 encoder 63 transmission line 64 decoder 65 D / A converter unit 66 the audio signal output terminal 67 the changeover switch 68 Attenuator

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI technical display location H04B 14/04 Z 4101-5K

Claims (2)

[Claims] 1. A silent section detecting means for detecting a silent section of an input voice signal, an input spectrum pattern measuring means for measuring a spectrum pattern of the input voice signal, a noise spectrum pattern group prepared in advance, and the noise spectrum. A neural network that has learned voice signal extraction for each noise spectrum pattern in the pattern group, and an encoding means are provided, and the spectrum pattern of the input voice signal at that time and the noise spectrum pattern at each timing detection of the silent section The closest noise spectrum pattern is selected by comparing with the individual noise spectrum patterns in the group, and the input speech signal converted into the frequency domain is input to the neural network learned in advance with the selected noise spectrum pattern. , Of the audio signal Speech coding scheme and performing coding after the extraction. 2. A silent section detecting means for detecting a silent section of an input voice signal, an input spectrum pattern measuring means for measuring a spectrum pattern of the input voice signal, a time measuring means, and a constant controlled by the time measuring means. A time zone spectrum pattern accumulating means for accumulating a time zone spectrum pattern, which is a spectrum pattern at the time of no sound, a comparison voice signal recorded in a noiseless state, a neural network group having a learning function, and an encoding means A voice signal is extracted by the sum of the time-range spectrum pattern and the comparison voice signal, and learning is performed in advance, and only the voice signal is adaptively extracted for noise that changes depending on the time zone. A voice encoding method characterized by performing encoding.