KR20010066558A - Voice activity detection method of voice signal processing coder using energy and LSP parameter - Google Patents

Abstract

PURPOSE: A method for voice activity detection in voice signal process encoder is provided to reduce a transmission rate at a mute interval in a voice signal process encoder by using energy and an LSP(line service protocol) parameter. CONSTITUTION: A first process(S10) calculates an average energy of a frame on the ground of a voice activity detection. A second process(S20) compares the calculated average energy with a noise level, and makes a decision of a voiced sound if the average energy is higher than the noise level and makes a decision of a voiceless or mute sound if not. A third process(S30) judges by using a minimum value and a maximum value of an LSP interval to consider a case that SNR(signal to noise ratio) is low if the second process(S20) makes a decision of a voiced sound. A fourth process(S40) compares the minimum interval with the maximum interval of the LSP to consider a case that the energy of voice is low if the average energy is lower than the noise level in the second process(S20).

Description

Voice activity detection method of voice signal processing coder using energy and LSP parameter

The present invention relates to a voice activity detection method using energy and LSP parameters in a method of lowering a transmission rate in a silent section in a voice signal processing encoder.

In the speech signal processing encoder, a voice activity detector and a comfort noise generator are used as a method of lowering the transmission rate in the silent section. In the VAD, the presence or absence of a voice signal in a frame is set to 1, otherwise, it is set to 0 to set the operation of the CNG.

However, the VAD of the G.723.1 vocoder has various conditions for the continuity and stability of the decision, and the condition determines whether there is a voice signal. First, the pitch characteristics of voiced sound discrimination and sinusoidal wave are detected, and the characteristics of the spectrum are used for accurate discrimination of signals with low signal-to-noise ratio.

The conventional VAD uses an adaptation enable flag for stability and continuity of discrimination. Since this value uses a characteristic of slow increase and fast decrease, silence exists when the silence section does not continuously occur. The interval is set to the frame in which the voice exists. The biggest problem with VAD is that it must be able to detect speech signals against any background noise. For example, even in a signal with a very low signal-to-noise ratio, it is necessary to accurately determine the presence or absence of a voice signal. However, this determination has a problem that it is almost impossible to distinguish between the speech and noise signals when the noise signal is lower than the speech signal as a simple discrimination condition.

The present invention has been made to solve the problems of the prior art, and an object of the present invention is to consider the characteristics of the spectrum in order to determine the presence or absence of a speech signal, so in VAD, an inverse filter using coefficients generated in a noise section is used. For the sake of stability and continuous discrimination, at the beginning of the signal, it is set to 1 in almost all cases, so that the LSP is easier to detect than the conventional method within the range that does not impair the stability. The present invention provides a method for determining the presence or absence of speech using a parameter.

As a technical idea for achieving the object of the present invention, the first step of calculating the average energy for the frame by Voice Activity Detection, the average energy is compared by comparing the calculated average energy and noise level If it is greater than the noise level, the voiced sound is determined. Otherwise, the second process is determined as the unvoiced or silent section, and if the voiced sound is determined in the above process, the minimum and maximum values of the LSP interval are considered to consider the case where the signal-to-noise ratio (SNR) is low. A third process of determining using a value and a fourth process of comparing the maximum and minimum intervals of the LSP to consider the case where the energy of the voice is small when the average energy is smaller than the noise level. The invention is presented.

1 is a block diagram showing a speech signal processing encoder having a speech activity detection device and a comfortable noise generator.

2 is a block diagram illustrating an example of a voice activity detector.

3 is a flowchart illustrating a voice activity detection algorithm using energy and LSP parameters according to the present invention.

<Brief description of the main parts of the drawing>

1: Input 2: Analog-to-Digital Converter (ADC)

3: inverse filter analyzer 4: autocorrelation coefficient (ACF)

4a: buffer 4b: average

5: weight adder 6: adder

7: comparator 8: voice / non-voice output

14: autocorrelation coefficient 15: buffer

20: control signal generation circuit 21: LPC analyzer

24: Buffer 27: Pitch Analyzer

Hereinafter, with reference to the accompanying drawings with respect to the configuration and operation of the embodiment of the present invention will be described.

1 is a block diagram illustrating a speech signal processing encoder having a speech activity detection device and a comfort noise generator.

As shown in Fig. 1, the encoding stage is composed of a speech signal processing encoder, a voice activity detector (VAD), a coding side comfort noise generator (COD-CNG), and a multiplexing device (MUX).

The voice activity detection apparatus (VAD) determines whether a voice is present for each 30 ms frame generated by the voice encoder. The VAD decision in frame t is denoted by Vad _t , which is the input to the COD-CNG block that computes Ftype _t .

2 is a block diagram illustrating an example of a voice activity detector.

As shown in Fig. 2, the voice activity detector includes an analog to digital converter (2), an inverse filter analyzer (3), an ACF (4), a buffer (4a), an AV (4b), a coefficient multiplier (5), and an adder (6). ), A comparator 7, a control signal generation circuit 20, and the like.

The input signal 1 is sampled by the analog-to-digital converter 2, digitized and input to the inverse filter analyzer 3. The inverse filter analyzer 3 is a part of the speech coder in which the actual speech activity detector operates and produces coefficients of the filter corresponding to the inverse of the spectrum of the input signal. The digitized signal is supplied to an autocorrelator 4. The autocorrelator 4 creates an autocorrelation vector of the input signal. The autocorrelation coefficients are averaged over several consecutive speech frames to improve reliability. This is because the autocorrelation coefficient output by the autocorrelator 4 is stored in the buffer 4a and the weighted sum of the autocorrelation coefficients of the previously stored frame provided by the averaging member 4b from the buffer 4a. By obtaining The average autocorrelation coefficients are transmitted to weighter 5 and adder 6. The weighter 5 and the adder 6 also receive autocorrelation vectors of the inverse filter coefficients of the noise period stored from the autocorrelator 14 via the buffer 15. A measure M is made from the mean autocorrelation coefficient and the autocorrelation vector. The measure is the threshold of the comparator 7. The logic result of the comparator indicates the presence or absence of speech at the output 8.

It is desirable to update the coefficients during the noise so that the inverse filter coefficients correspond to the inverse predictions of the noise spectrum. However, the determination of voice / non-voice based on update does not depend on the result of the update. One misidentified frame of the signal is followed by the voice activity detector incorrectly recognizing the next frame. Therefore, there is a control signal generation circuit 20, so that the inverse filter autocorrelation coefficient used to make the measure M by making an independent control signal indicating the presence or absence of speech to control the inverse filter analyzer 3 is updated only during the noise period. The control signal generation circuit 20 includes an LPC analyzer 21 for producing LPC coefficients corresponding to the input signal and an autocorrelator 21a for generating autocorrelation coefficients. The autocorrelation coefficients are supplied from the autocorrelator 4 to the weighter 22 and the adder 23 which receive the autocorrelation vector of the input signal. The degree of spectral similarity between the input speech frame and the previous speech frame is calculated. The difference signal on the spectrum determined by the comparator 26 indicates the presence or absence of voice. Although such a measurement is excellent in distinguishing noise from non-voice, a speech detection circuit including a pitch analyzer 27 is provided in the circuit 20 because it is insufficient in distinguishing noise and speech. The pitch analyzer 27 produces a logical signal that is true when a voice signal is detected, and this signal is a "true" signal when the voice is present along with a threshold measure from the comparator 26. It is input to the generated NOR gate 28. The signal is supplied to the buffer 15 so that the inverse filter coefficients are updated only during the noise period.

The threshold adapter 29 is connected to receive the non-voice signal control output of the control signal generation circuit 20. The output of the threshold adapter 29 is supplied to the comparator 7. The threshold adapter 29 incrementally increases or decreases the threshold until the threshold approximates the noise level. When the input signal is very small, it is desirable that the threshold is automatically set at a fixed low level. Because at low signal levels, the effect of signal quantization by the ADC 2 may lead to undesirable results.

3 is a flowchart illustrating a voice activity detection algorithm using energy and LSP parameters according to the present invention.

As shown in FIG. 3, a first step S10 of calculating an average energy for a frame by Voice Activity Detection, and comparing the calculated average energy with a noise level, the average energy is a noise level. If it is greater than the second step (S20) is determined as a voiced sound, otherwise determined as an unvoiced or silent section, and the minimum value of the LSP interval to consider the case where the signal-to-noise ratio (SNR) is low when it is determined as a voiced sound in the process The third step (S30) to determine by using and the maximum value, and the second step is determined to compare the maximum interval and the minimum interval of the LSP to consider the case where the energy of the voice is small when the average energy is less than the noise level It is composed of a fourth process (S40).

In the third process, if the LSP minimum interval is greater than 1/2 of the maximum interval, the formant is set to voice activity detection (S31). And increasing the level (S32), wherein the fourth process sets the case where the voice is present when the LSP minimum interval is less than 1/2 of the maximum interval and reduces the noise level (S41); If not, it is a step (S42) for determining as unvoiced or silent.

First, VAD calculates the average energy for a frame with 240 samples. Compare the average energy and noise level. At this time, if the average energy is greater than the noise level, it is judged as voiced sound. Otherwise, it is judged as unvoiced or silent section. In the case of voiced sound, the final decision is made by using the minimum and maximum values of the LSP interval to consider the case where the SNR is low. If the LSP minimum interval is greater than half of the maximum interval, the formant is present and VAD = 1 is set. Otherwise, the energy of the noise is determined to be a large signal and the level of noise is increased.

If the average energy is smaller than the noise level, the maximum and minimum intervals of the LSP are compared in a similar manner as above to consider the case where the energy of the voice is small. If the LSP minimum interval is less than 1/2 of the maximum interval, since voice is present, VAD = 1 is set and the noise level is reduced.

VAD is basically an energy detector. The energy of the inverse filtered signal is compared with a threshold and if it exceeds this threshold it is determined that voice is present in the frame. Two steps are required to calculate the threshold. First, the noise level is updated based on the noise level of the previous frame and the inverse filtered energy of the current frame. Second, the threshold is calculated using the log scale noise level.

An adaptation enable flag, denoted Aen _t for the current frame t, is used to ensure that the VAD noise level is updated only in the absence of speech. This is based on the fact that background noise is neither a speech signal nor a sinusoid.

The open loop pitch delay of the previous and current frames is used to determine the voiced sound.

When this value is L ^j _OL , j = 0,1,2,3, the minimum delay value L ^min _OL = Min (L ^j _OL , j = 0,1,2,3) is calculated first. Counter pc ∈ [1,2,3,4] is intended to indicate how much delay L ^j _OL is around the L ^min _OL (± 3) multiple. If pc is 4, the signal is determined to be voiced.

The following sinusoidal detector is included in the LPC analysis of the G.723.1 encoder.

^t _i k [2] for each subframe of the frame t i = 0, ..., when considered with the second reflection coefficient calculated by the dyubin (Durbin) algorithm if the at least three 14 values are k ^t from among the 15 values _{If i} [2]? 0.95, it is determined that a sine wave is detected (SinD = 1). Otherwise, SinD = 0.

The equation for calculating the adaptive enable flag is as follows.

Aen _t = Aen _t-1 +2 if pc = 4 or SinD = 1

Aen _t = Aen _t-1 -1 otherwise

In Equation 1, Aen _t is a boundary condition of [0,6].

The input signal frame, {s [n]} _{n = 60..239} , is inversely filtered by the FIR filter A _no (z) with the coefficient {a _no [j]} _{j = 1..10} . This filter is calculated by the CNG block and provides an LPC filter that is related to the background noise of the current frame.

In Equation 2, e ' _t is an inverse filtered signal.

The energy, Enr _t, is calculated from the inverse filtered signal of the current frame.

The noise level of frame t, Nlev _t , is updated by the previous noise level and the previous energy, Enr _t-1 , and the adaptive enable flag, Aen _t . This update process is characterized by a slow increase, a fast decrease. The dynamic range of the noise level in frame t is limited to [Nlev _min , Nlev _max ].

If Nlev _t-1 > Enr _{t-1, the} noise level is clipped.

If adaptation is active, Nlev _t is increased, otherwise it is decreased slightly.

with

The relationship between the noise level, Nlev _t , threshold, Thr, in frame t is defined in logarithmic scale and uses the following formula:

The VAD decision is determined by comparing the threshold Thr with the current energy Enr _t .

All static variables of the VAD algorithm are initialized to zero except for the variables in the following formula.

Nlev _-1 = 1024

Enr _-1 = 1024

L _OL ^j = 1 j = 0,1

L _OL ^j = 60 j = 2,3

In the voiced spectrum, the first formant generally appears within 1 kHz and three or more formants exist. On the other hand, the unvoiced spectrum has low energy in the low frequency region and high energy in the high frequency region. Like this spectral shape, the shape of the LSP also differs significantly between voiced and unvoiced sounds. First, in voiced sound, more LSPs are distributed in the low frequency region due to the formant, and the interval is narrower than that of the high frequency LSP. On the other hand, in the unvoiced sound, the LSP is distributed in the high frequency region and the spacing is narrower than that in the low frequency region.

Due to the characteristics of the LSP parameter according to the voiced and unvoiced sounds, component separation is possible. First, the difference of distribution according to LSP voiced and unvoiced sound is used. If the sampling frequency is Fs, and the number of LSPs in the frequency domain below Fs / 4 is NL, and the number of LSPs in the frequency domain above Fs / 4 is NH, when NL is larger than NH, It is considered that the spectrum is characterized by the appearance of many poles on the low frequency side, indicating the spectral characteristics of voiced sound. That is, the first formant and the second formant of the voiced sound mainly exist in the low frequency region.

In contrast, if NH is greater than NL, it is determined that it is silent. In other words, the unvoiced spectrum is because the main formant appears in the high frequency region. However, voiced sound such as / i / I / ε / ae / has a second formant, a third formant, or a fourth formant in the high frequency region, so that NH is larger than NL even though it is a voiced sound. In such a case, whether the first formant is present or not is determined whether it is an unvoiced sound or a voiced sound. In other words, if there are LSPs with narrow intervals in the region below Fs / 4 by examining the intervals of the LSP parameters, they are regarded as voiced sounds.

As can be seen from the above description, the present invention can reduce the transmission rate by making the detection of the silent section more frequently than the conventional G.723.1 voice activity detection device, and thereby obtain the user increase effect. have. In addition, the present invention can be used as an algorithm for detecting silent sections as well as voice activity detection and speech recognition, and effective results can be obtained when detecting voiced / unvoiced sections. The LSP minimum interval information is required for speech analysis, and this information indicates the position of the first formant. Therefore, it is possible to determine whether the voice / voice is sound by adding the algorithm proposed in the present invention.

Claims (3)

In the method of lowering the transmission rate in the silent section in the speech signal processing encoder, A first process of calculating an average energy for a frame by Voice Activity Detection, A second process of comparing the calculated average energy with the noise level and determining that the average energy is greater than the noise level as voiced sound; A third step of determining by using the minimum and maximum values of the LSP interval in order to consider the case where the signal-to-noise ratio (SNR) is low when it is determined as voiced sound in the above process; Speech signal processing encoder using energy and LSP parameters, characterized in that the fourth step of comparing the maximum and minimum intervals of the LSP to determine the case where the average energy is less than the noise level is compared to determine the maximum and minimum intervals of the LSP Voice Activity Detection in. The method of claim 1, wherein the third process comprises: setting the voice activity detection that the formant exists when the LSP minimum interval is greater than 1/2 of the maximum interval; Otherwise, determining the signal with a large energy of noise and increasing the level of noise. The method of detecting a voice activity in a speech signal processing encoder using energy and LSP parameters. The method of claim 1, wherein the fourth process comprises the steps of setting the case where the voice is present when the LSP minimum interval is less than 1/2 of the maximum interval and reducing the noise level; Otherwise, the voice activity detection method in the speech signal processing encoder using the energy and the LSP parameter, characterized in that it is determined as unvoiced or silent.