JPH1124692A - Method of judging generating/resting section of voice wave and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To minimize the effect by background noise and to minimize the operation quantity. SOLUTION: An input voice is power-detected in a short time (31), and supplied to a neural network(NNW) 34 directly or after sequentially delayed by a plurality of delay elements, and the maximum self-correlation of the input voice is detected in a range longer than the pitch period (32), and inputted to the NNW 34 to determine an LSP(linear spectrum pair) coefficient of the input voice (33). The error of this from a specified LSP vector is calculated (35) and inputted to the NNW 34. The NNW 34 is allowed to preliminarily learn to output 1 in sounding and 0 in resting section, and the output of the NNW 34 is threshold-judged (37) to provide a judgment result of sounding/resting section.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applicable to the field of digital transmission of voice and the like, and belongs to the field of digital processing of voice, and a method and apparatus for discriminating between a sound section and a pause section in a voice wave. About.

[0002]

2. Description of the Related Art First, a sound section and a pause section in a speech waveform will be described. FIG. 9 shows an example of a sound section and a pause section in an actual speech waveform. In FIG. 9A, it is clear that pause periods T _US and T _UE exist due to the fact that no voice is uttered before and after the voice. In addition, it can be seen that a very short pause section T _UM exists even while the voice is being uttered. FIG. 9B is an enlarged view of the vicinity of the pause section T _UM existing in the utterance section. Pause section T seen in FIG.
_UM appears immediately before a pop or fricative.

[0003] In these pause sections T _US , T _UE and T _UM , the speech wave has no power, so in voice information transmission, it is sufficient to transmit only information that these sections are pause sections. A method for reducing the total amount of transmitted information by transmitting a pause section with a very small amount of information for the purpose of effectively utilizing the frequency resources of the transmission path in the field of voice digital transmission. Is used.

[0004] In order to realize such a voice transmission method, a voiced / paused section determination device for a voice which can accurately specify whether a voice is a voiced section or a paused section is required. The simplest method of determining a sound section and a pause section of a voice is a method of measuring a short-time power of a voice wave and comparing the measured power with a certain threshold. However, this method is susceptible to voice level fluctuations and background noise, and is liable to cause erroneous voice / pause determination.

[0005] As a method of solving such a problem, there is a document "A study of a sound / non-sound detection circuit in VOX control", IEICE Spring Conference, B-422, 19
There are 93. FIG. 10 shows the configuration. The input sound 11 is processed by the bandpass filter 12, amplified by the amplifier 13,
The data is PCM-coded by the PCM encoder 14 and output at the output terminal 15. The power of the PCM-coded speech is calculated by a silence detection circuit 16 and compared with a predetermined threshold value to determine a sound / pause section. On the other hand, the level of the output of the amplifier 13 is detected by the detection circuit 17 and transmitted to the gain control circuit 19 via the hold circuit 18 at the detected level.
3 is adaptively controlled. Furthermore, the silence detection circuit 1
6 controls the hold circuit 18 so as to hold the output of the detection circuit 17 while the idle period is being detected, and operates so as not to perform the gain control in the idle period.

In this method, the automatic gain control circuit 1
9 controls the gain of the amplifier 13 to make the output of the amplifier 13 a certain level, thereby reducing the erroneous determination of the sound / pause section caused by the level fluctuation of the input voice and stopping the silence detection circuit 16 While the section is detected, the operation of the automatic gain control circuit 19 is held to prevent the gain of the amplifier 13 from becoming unnecessarily large in the voice pause section, and to determine a sound / pause section caused by background noise. The error has been suppressed.

[0007] As another method, there is a method of determining a sound / pause section using an autocorrelation coefficient or a pitch lag of an audio wave in addition to the short-time power of the audio wave. As an example of such a device, the document “Europian digital cellular te”
lecommunications system (Phase 2); Voice Activity D
etection (VAD) (GSM 06.32) ", European Tel
There is the ecommunications Standards Institute (1994).
FIG. 11 shows the configuration. After the input speech is subjected to PCM encoding, an autocorrelation coefficient and a pitch lag are analyzed, and the speech is input to the apparatus from an autocorrelation coefficient input terminal 21 and a pitch lag input terminal 22, respectively. The autocorrelation coefficient is sent to the residual power calculator 23, and the linear prediction residual power of the section to be subjected to the sound / pause section determination is calculated. At the same time, the auto-correlation coefficient is sent to the auto-correlation coefficient averaging unit 24,
The average value of the autocorrelation coefficients for the past several frames including the section to be subjected to the pause section determination is calculated, and the spectrum comparison unit 2
5 and the past average value of the autocorrelation coefficient is sent to the predicted value calculating unit 26. The predicted value calculating section 26 predicts the current average value of the autocorrelation coefficient using the past average value of the autocorrelation coefficient, and sends it to the spectrum comparing section 25. The spectrum comparing unit 25 compares the spectra from the two averaged autocorrelation values and determines the continuity of the spectrum. The input pitch lag is sent to the periodicity determination unit 27, where the continuity of the pitch is determined. The residual power, the spectrum continuity determination result, and the pitch continuity determination result are sent to the threshold value adaptation unit 28, where the threshold value of the residual power for sound / pause section determination is determined. The determined threshold value is compared with the residual power by the VAD determination section 29, and a sound / pause section is determined.

FIG. 12 shows the processing of the threshold adaptation unit.
If the residual power is smaller than the tentative determination threshold value pth (constant) (S1), it is unconditionally determined to be a pause section, the sound / pause section determination threshold value thvad is set to the initial value plev, and the processing ends. (S2). If the residual power is equal to or greater than pth (S1) and the spectrum is unsteady (S3) or the pitch is steady (S4), the sound section is unconditionally determined, and the change of the sound / pause section determination threshold is performed. The processing is not performed, and the value count of the threshold adaptation number counter is set to 0, and the process ends (S5).

In other cases, that is, when the residual power is pth
As described above, when the spectrum is stationary and the pitch is not stationary, it is treated as a background noise section, and when this continues more than a certain number of times adp, the sound / pause section determination threshold is changed. That is, when it is determined that the background noise period is present, the count value of the threshold adaptation number counter is incremented by +
1 (S6), and the counted value count becomes the adaptive grace number a
If it does not exceed dp, the processing is terminated (S7). If it does, the sound / pause section determination threshold value thvad is divided by the threshold value change step size coefficient dec, and the result is subtracted from the threshold value thvad. A new threshold value thvad is set (S8). Next, the result of multiplying the residual power pvad by the audio power / residual power comparison coefficient fac is a threshold th.
If it is larger than vad (S9), the threshold value thva
The value obtained by adding thvad to a value obtained by dividing d by a coefficient inc defining the lower limit of the threshold value change, and a residual power pva
The smaller of the product of d and the product of the audio power / residual power comparison coefficient fac is set as a threshold thvad (S10), and the threshold thvad or the threshold th in step S9.
If the threshold value thvad when vad does not exceed pvad × fac is larger than the sum of the residual power pvad and the residual power margin margin, then add the residual power pvad and the margin margin. After the threshold value thvad is set (S12), or when it is determined in step S11 that the threshold value thvad is not larger, the adaptation grace number adp is incremented by 1 to be the count value count of the threshold adaptation number counter. End (S1
3).

In this method, not only the power of the voice but also information such as the continuity of the spectrum and the continuity of the pitch is used to adaptively change the threshold value of the power for the determination of the sound / pause section. As a result, it is possible to reduce the determination error in the voiced / pause section determination due to the voice level fluctuation and the background noise.

[0011]

The conventional method using the automatic gain control circuit has a problem in that the input audio wave is amplified by using an amplifier having a variable gain, so that the audio wave can be faithfully transmitted. On the other hand, in the method using the autocorrelation coefficient and the pitch lag, a large number of controls such as comparison with a threshold value and conditional branching are required due to its configuration, and there is a problem that a control circuit is complicated in realizing the device.

[0012]

According to the first aspect of the present invention,
One or more short-time speech powers, a parameter vector related to a spectrum envelope obtained by analyzing the speech wave, a vector-to-vector distance between a parameter vector having the same kind of parameter and a flat spectrum, and speech. And the maximum value of the autocorrelation of the sound wave within a range substantially covering the pitch period of the sound wave is input to the neural network, and the output of the neural network determines the sound / pause section.

According to a second aspect of the present invention, the variance of the quantized sampling value of the audio wave in one or a plurality of short time intervals is used instead of the short-time audio power in the first aspect of the present invention. Are different. In each of the first and second aspects of the present invention, since it is not necessary to amplify the input voice using an amplifier having a variable gain, it is superior to the method using an automatic gain control circuit in terms of faithful transmission of a voice wave. .

In each of the first and second aspects of the present invention, the sound / pause section is determined using a plurality of parameters in the same manner as in the method using the autocorrelation coefficient and the pitch lag. Since the sound / pause section is determined using only the output obtained by inputting to the network, there is an advantage that only one comparison with the threshold value is required and no conditional branch control is required.

Further, according to the second aspect of the present invention, even when the sound wave has a bias, for example, when the zero level of the AD converter and the zero level of the microphone input do not match, the sound / voice is stably output. It is possible to determine a pause section.

[0016]

FIG. 1 shows an embodiment of the invention described in claim 1. The input voice from the terminal 11 is sent to the short-time voice power calculation unit 31, the maximum autocorrelation calculation unit 32, and the LSP coefficient calculation unit 33, where the short-time voice power, the maximum autocorrelation, and the line spectrum pair (LSP) parameter are calculated. Is done. The calculated short-time voice power is directly sent to the neural network unit 34, and is sequentially passed through the delay elements 35 ₁ , 35 ₂ ,..., 35 _n and the outputs of the respective delay elements are respectively sent to the neural network unit 31. Sent.

The maximum autocorrelation calculating section 32 calculates an autocorrelation coefficient up to a time delay sufficient to reproduce the pitch period of the input voice, and the maximum value is calculated by the neural network section 3.
4 is input. L calculated by the LSP coefficient calculator 33
The SP parameter vector is an LSP vector error calculator 3
LS of the flat spectral envelope preset in 6
A vector error with the P parameter vector is calculated, and the obtained vector error is input to the neural network unit 34.

The neural network section 34 is provided with a large number of learning voices together with information on voiced / paused sections, and is trained to output, for example, 1 for voiced sections and 0 for silent sections. For this learning, a general method such as an error back propagation method may be used. The output of the neural network unit 34 is compared with a certain threshold value in the sound / pause section determination unit 37 to determine the sound / pause section.

The short-time voice power calculating section 31 uses the same method as that of the conventional method for detecting a sound / pause section of this type, and has a similar time section, for example, 5 to 20 ms.
The audio power is calculated for each degree. The maximum autocorrelation calculation unit 32 finds the maximum value of the autocorrelation within a range substantially covering the voice pitch interval. Delay element 35
Each delay time of ₁ , 35 ₂ ,... Is made equal to the calculation time section of the short-time voice power. That is, for example, when calculating the short-time voice power every 5 ms, the delay elements 35 ₁ ,.
Each delay time of 35 ₂ ,... Is 5 ms. It is preferable that the shortest audio power input to the neural network unit 34 has the longest delay of about 15 to 40 ms. That is, if the calculation time section of the short-time audio power is 5 ms, about 3 to 10 delay elements are preferable. If the delay time of the capture is short, the determination performance is reduced, and it is particularly preferable to be about 20 ms. Even if the delay time is longer than this, the determination performance is not so good and the processing amount is increased.

The sound / pause section is determined, for example, by 10 to 4
It is performed for each fixed time section of about 0 ms, and the calculation time section of the short-time audio power is selected to be equal to or shorter than the fixed time section for this determination. According to this configuration, not only the short-time audio power is used as one factor of the determination as in the conventional case, but also the use of the maximum value of the autocorrelation allows the sound to be determined as a sound if the pitch is stationary. The distance has a bias in the frequency characteristic of the spectrum.
In other words, if the characteristics are not flat, it can be determined that there is sound (the spectrum of the noise is flat). Using these multiple parameters, similar to the conventional method shown in FIGS. It can be judged correctly without being affected.

Since the LSP vector error calculating section 36 is for detecting the deviation of the spectrum as described above, not only the LSP parameters but also the LPC (linear prediction coefficient), the PARCOR coefficient, etc. Any parameters can be used as long as the parameters can be shifted. The invention according to claim 2 is implemented by replacing the short-time power output unit 31 with the short-time variance calculation unit in the invention according to claim 1. In the short-time variance calculation, the variance of the quantized sample values in the short-time section is calculated. The variance is large in the case of speech, and small in the case of noise. Also, if the input has a bias, such as when the zero level of the AD converter and the zero level of the microphone input do not match,
Although a relatively large short-time audio power is detected in the pause section, the sample value is a fixed value, so that the variance becomes extremely small, is distinguished from a sound, and there is no possibility of erroneous determination.

In the first and second aspects of the present invention, only the short-time power calculating section and the short-time variance calculating section are different from each other. FIG. 2 shows the analysis conditions of each parameter and the like. The input voice is sampled at a sampling frequency of 8000 [Hz], and the voice power is calculated every 5 [ms]. The autocorrelation coefficient is calculated for all the sample delays in the range shown in FIG. 2, and the maximum autocorrelation value is obtained.

FIG. 3 shows an LSP parameter vector as a reference for calculating an LSP vector error. From this figure, the intervals of each order are equal, which indicates that the spectrum is flat. The LSP vector error is the LSP vector shown in FIG.
It was defined as the Euclidean distance of the SP vector. The neural network unit 34 uses a four-layer model, and the number of neurons in the input layer, the first intermediate layer, the second intermediate layer, and the output layer is set to 7, 3, 3, and 1, respectively. The output o _{i, j} of the j-th neuron in the i-th layer is o _{i, j} = (1 / (e− ^x + 1) − (1/2)) (1) where x = Σ _k w _{ik, j} o _{i -1, k + β i, j} (2) here, w _{ij, k} is a weighting coefficient when the output of the i-1 layer k-th neuron is input to the i-th layer j-th neuron, also beta _{i , j} is a bias for the input of the i-th layer neuron j. However, the input layer is the 0th layer, and the output layer is the 3rd layer. Since equation (1) takes an output in the range of -0.5 to 0.5, 0.5 is added to the output of the output layer to make the output range 0 to 1. FIG. 4 shows an example of each weight coefficient, and FIG. 5 shows an example of each bias.

According to a second aspect of the present invention, in the first aspect of the present invention, the short-time voice power calculating section is changed to a short-time variance calculating section, and the weight and bias of the neural network section 34 are changed as shown in FIGS. It can be realized by changing to

[0025]

According to the first and second aspects of the present invention, in addition to the short-term power of speech, a vector error of a parameter vector related to a spectral envelope with a vector of its flat characteristic and a self-wave of a speech wave within a predetermined time are obtained. The use of the maximum correlation value has the effect of reducing the effect of background noise on voiced / pause section determination.

Further, by processing a plurality of parameters by using a neural network, conditional branch processing is not required at all, which has the effect that the device can be realized without using a complicated control circuit.
Comparing the required calculation amount of the present invention with the method using the autocorrelation coefficient and the pitch lag of the conventional method, the present invention operates with about half the calculation of the conventional method when using the parameters shown in the embodiment. It is estimated that.

Further, in the invention according to the second aspect, the short-time dispersion of the instantaneous value of the sound wave is used without using the short-time power of the sound, so that even when a constant bias is superimposed on the sound wave, it is stable. This makes it possible to determine a sound / pause section. FIG. 8 shows a comparison between the conventional method of making a sound / pause section judgment by comparing the short-time power of speech with a certain threshold value and the judgment test result according to the present invention. In the same judgment test, a sound in a state without background noise is used as a reference based on a result of judgment of a sound / pause section by a method (conventional method) based on comparison with a short-time power threshold value. At this time, the threshold value is -45 [dB] compared to the long-term average power of the voice.

When the present invention is applied to a speech without background noise (shown as noise-free in the figure), there is a slight error in determining a voiced section as a pause section, but most of the errors are large. A sound other than a sound having power, for example, a sound such as a breathing sound or a tongue tapping, is determined to be a pause section and does not cause any particular problem. As a result of performing a sound / pause section determination on a voice on which background noise of -20 [dB] is superimposed on the long-term average power of the voice,
In the conventional method, the threshold values are -45 [dB] and -30.
In the case of [dB], no pause section was determined.
Also, by setting the threshold to -20 [dB], it was possible to correctly determine about 60% of the pause section. On the other hand, when the present invention was used, about 95% of the pause section was correctly determined. The content of the determination error of the sound section at this time was the same as the case where the background noise was not superimposed.

-20 compared to the long-term average power of voice
[DB] Voice with superimposed bias
As a result of performing the pause section determination, when the invention according to claim 1 is used, the pause section can hardly be determined. However, when the invention according to claim 2 is used, about 65% of the pause section can be correctly determined. did it.

[Brief description of the drawings]

FIG. 1 is a diagram showing a functional configuration according to an embodiment of the present invention;

FIG. 2 is a diagram showing a numerical example of each parameter when the invention of claim 1 is implemented.

FIG. 3 is a diagram showing an example of a reference LSP parameter vector.

FIG. 4 is a diagram showing an example of each weight coefficient of the neural network unit 34 in the embodiment of the first invention.

FIG. 5 is a diagram showing an example of a bias β _{i, j} of the neural network unit in the embodiment of the first invention.

FIG. 6 is a diagram showing an example of each weight coefficient of the neural network unit 34 in the embodiment of the second invention.

FIG. 7 shows a neural network according to the second embodiment of the present invention.
Bias of the network section 34 _{i, j}FIG.

FIG. 8 is a diagram showing a result of performing a sound / pause section determination by the method of the present invention and the conventional method.

FIG. 9 is a waveform diagram showing an example of a sound section and a pause section in an audio wave.

FIG. 10 is a diagram showing a functional configuration of a sound / pause section determination device using conventional automatic gain control.

FIG. 11 is a diagram showing a functional configuration of a conventional voiced / pause section determination device for voice using an autocorrelation coefficient and a pitch lag.

FIG. 12 is a flowchart showing a procedure of a threshold adaptation process in the conventional device of FIG. 11;

Claims (4)

[Claims] 1. A speech waveform sampled and quantized at a fixed cycle, divided into fixed time sections, and each time section contains a sound section or a pause section. In the method of determining whether or not a parameter vector related to the spectral envelope is obtained by analyzing the speech waveform, the vector distance between the parameter vector and a parameter vector of the same kind and having a substantially flat spectral envelope is determined. Determining a short-time audio power of the audio waveform; obtaining a maximum value within a range substantially covering a pitch period of the audio of the auto-correlation of the audio waveform; determining the vector distance and at least one of the short-time audio powers; , And the maximum value of the autocorrelation is input to the neural network, and one output thereof is compared with a threshold value.
A sound / pause section determination method for a speech wave, which determines whether the voice waveform is a voice section or a pause section based on the magnitude. 2. A sound waveform sampled and quantized at a fixed cycle and divided into fixed time sections, and for each time section, the sound contained therein is a sound section or a pause section. In the method of determining whether or not a parameter vector related to the spectral envelope is obtained by analyzing the speech waveform, the vector distance between the parameter vector and a parameter vector of the same kind and having a substantially flat spectral envelope is determined. Determining a variance of a quantized sampling value of the audio waveform in a short time period of the audio waveform, obtaining a maximum value within a range that substantially covers a pitch period of the audio of the autocorrelation of the audio waveform, A vector distance, at least one of the variances, and a maximum value of the autocorrelation are input to a neural network, and one output is output. Compared to have values, speech / pause interval determination method of speech waves and judging whether the pause interval the speech waveform is voiced section by its magnitude. 3. A device for inputting a sampled and quantized speech waveform at a fixed cycle and determining whether the speech contained therein is a voiced section or a paused section at regular time intervals, Means for analyzing the speech waveform to obtain a parameter vector which is not related to the spectrum envelope; and means for obtaining a vector distance between the parameter vector and a parameter vector of the same kind which has a substantially flat spectrum envelope. Means for obtaining a short-time audio power of the audio waveform; means for obtaining a maximum value within a range substantially covering a pitch period of the audio of the autocorrelation of the audio waveform; at least one of the vector distance and the short-time audio power A neural network which receives the maximum value of the autocorrelation and outputs an output from one output terminal; Characterized in that the output is compared with a threshold value to determine whether the voice waveform is a voiced section or a paused section based on the magnitude of the output. 4. An apparatus in which a speech waveform sampled and quantized at a fixed cycle is input and a speech interval included in the speech waveform is determined at regular time intervals to determine whether the speech is a speech interval or a pause interval. Means for analyzing the audio waveform to obtain a parameter vector related to the spectral envelope; and means for obtaining a vector distance between the parameter vector and a parameter vector of the same type that has a substantially flat spectral envelope. Means for obtaining a variance of a quantized sampling value of the audio waveform in a short time period of the audio waveform, and means for obtaining a maximum value within a range substantially covering a pitch period of the audio of the autocorrelation of the audio waveform, The vector distance, at least one of the variances, and the maximum value of the autocorrelation are input, and an output is output from one output terminal Means for comparing the output of the output terminal with a threshold value and determining whether the voice waveform is a voiced section or a paused section based on the magnitude thereof. Sound / pause section determination device.