JPH07135490A - Voice detector and vocoder having voice detector - Google Patents

Abstract

PURPOSE:To attain accurate voice detection with a simple circuit scale and less processing time by utilizing parameters calculated from a vocoder having an acoustic sense weighting filter adaptor and a backward gain adaptor efficiently. CONSTITUTION:A reflection coefficient (t) calculated by an acoustic sense weighting filter adaptor 106 of a vocoder 100 and an exciting gain (s) calculated by a backward gain adaptor 110 are inputted to a voice detector 200. The voice detector 2000 uses mean value calculation devices 201, 204 to calculate a mean value of an input signal for a predetermined period and a decision circuit 205 compares the mean value with a preset threshold level to make decision on sound/silence. Thus, accurate decision is made under a usual service environment by taking the presence of background noise in account. The quantity required for decision processing is less and the voice is detected for a period less than a 10ms frame unit and the circuit scale is made small.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder, and more particularly, to a speech detector for making speech / non-speech judgment of speech and a speech coder having the speech detector.

[0002]

2. Description of the Related Art The following can be considered as one application example of a voice detector. In a portable wireless device or the like, in order to reduce power consumption, VOX (Voice Opera) which transmits only when there is voice and suspends transmission when there is no voice
The rate switch exchange control is used, and if this is used, the average power consumption during transmission can be reduced by about 65% in full-duplex communication. In order to realize such a VOX function, it is necessary for the transmitting side to detect the presence or absence of an audio signal. An audio detector is provided to detect the presence or absence of this audio signal.

Description will be given below on the premise that this voice detector is applied to a VOX of a digital cordless telephone system. In this digital cordless telephone device, a processing delay time of 2 ms or less is required as a next-generation half-rate CODEC (voice coding method), and LD-CELP (low delay code excitation linear prediction) of 16 kbps is required. Is expected to be used. In addition, it is said that the frame length of voice / non-voice determination by the voice detector in this device is appropriate in 10 ms frame unit according to the data transmission format (see RCR STD-28 standard).

A block diagram of a conventional voice detector is shown in FIG. The voice detection here is 8 kHz sampling and 8
The bit-quantized input voice a is divided into 20 ms frame units (160 samples) and executed. In the direct-current component suppressor 11, a wide area filter is used to
Outputs the signal with the DC component removed.

In the high level power detector 12, 20 ms
Subframe every 32 ms (32 samples)
And the short-term power P _SK is calculated for each sub-frame by the formula 1.

[0006]

[Equation 1]

Furthermore, for each P _SK , the power threshold value Th2
By (-30dbm0), the weighting power D _{2K is} calculated by Equation _2.
Is detected.

[0008]

[Equation 2]

Next, the total weighted power D ₂ is calculated by the equation 3, and this is output as the detection result c of one frame.

[0010]

[Equation 3]

Further, in the low-level power detector 13, with respect to short-term power P _SK having 1, power闘値Th1 (-5
0dbm0) is used to detect the weighting power D _1K by Equation 4.

[0012]

[Equation 4]

Similarly, the weighted sum D ₁ is calculated by the equation 5, and this is output as the detection result d for one frame.

[0014]

[Equation 5]

At this time, D ₃ is also calculated by the following equation 6.

[0016]

[Equation 6]

In the zero-crossing number detector 14, the filter output b
In order to count the number of zero-crossings Z _SK (the number of sign bits between two consecutive audio signal samples inverted), the following equation 7 is calculated for each subframe.

[0018]

[Equation 7]

Next, for each Z _SK , the zero-cross threshold value Th3
(24), the zero-cross number DZ _{SK is} detected by the equation 8.

[0020]

[Equation 8]

Similarly, the weighted sum Dz is calculated by the equation 9, and this is output as the detection result e of one frame.

[0022]

[Equation 9]

Further, the inter-frame power increment comparator 15 obtains the power P _TN for one frame by the calculation of the following expression 10.

[0024]

[Equation 10]

By comparing with the inter-frame power P _{T (n-1)} of the previous frame, the power increment is detected by the following equation 11, and the result is output as f.

[0026]

[Equation 11]

The decision unit 16 inputs each signal, makes a decision by the decision logic flow shown in FIG. 7, and outputs a voice detection result g. In FIG. 7, HOT is a hangover timer (a function for forcibly determining the presence of speech in a few frames after that when the determination is changed from voiced to silence to prevent ending of the ending). The flag is a voice / non-voice flag.

In step 21, the high level power detector 12
Output C (D ₂ ) of the zero crossing number detector 14 in step 22.
Output e (Dz), the output d (D ₁₎ of the low-level power detector 13 at step 23, the output f (D ₄₎ of the frame power increment comparator 15 in step 24, and step 25
Then, each of the numerical values (D ₃ ) obtained by the low level power detector 13 is judged. Each judgment result follows the judgment theory flow, and when satisfied, SP flag set 27, HOTSET2
The voice flag is output from the SP flag transmission 31 via 8,
If not established, a silent flag is output.

[0029]

However, since the processing by the speech detector of the prior art is executed in the unit of 20 ms frame, a delay time of at least 20 ms occurs, and the condition of the unit of 10 ms frame cannot be satisfied.
Further, since this conventional speech detector is independent of the speech coder, it has the drawback of increasing the total processing amount or increasing the hardware scale.

An object of the present invention is to efficiently use the parameters obtained in the process of the speech coder having the auditory weighting filter adaptor and the backward gain adaptor, so that the processing time can be reduced and the delay time can be 2 ms. The speech coder is provided with a means for enabling speech detection in a unit of 10 ms frame or less that can be realized under the following speech coder, that is, a speech detector.

[0031]

In the speech coder having a perceptual weighting filter adaptor and a backward gain adaptor, the object is the inverse coefficient calculated by the perceptual weighting filter adaptor and the backward gain adaptation. This is accomplished by providing the speech encoder with a speech detector that makes a voiced / silent decision by inputting each of the excitation gains calculated by the device and comparing it with a prepared threshold value.

[0032]

According to the present invention, the auditory weighting filter adaptor of the encoder outputs the reflection coefficient t by the internal calculation of the input signal vector. This is a vocal tract parameter that indicates the spectral envelope information of the input signal.
It is possible to judge voice and background noise. The backward gain adaptor inputs the excitation signal whose gain has been adjusted, and outputs the excitation gain s by backward adaptation. This can be considered as the input signal power, and the voice can be detected by the power judgment based on this parameter. The speech detector has a reflection coefficient t and an excitation gain s output from the encoder.
Is input, and a voice detection signal is output by comparing these two types of parameters with a threshold value prepared in advance, and it is possible to determine whether there is sound or no sound.

[0033]

EXAMPLES The present invention will be described below with reference to examples. 16 kbps LD- expected to be used as a next-generation half-rate voice encoder for digital cordless telephone devices
An example applied to CELP is shown below.

FIG. 1 shows an LD-CEL having a voice detection function.
1 shows a P speech encoder 100. In FIG. 1, 101 is a uniform PCM conversion circuit, 102 is a vector buffer, and 10 is a vector buffer.
3 is an excitation codebook, 104 is a gain adjustment unit, 1
Reference numeral 05 is a synthesis filter, 106 is a perceptual weighting filter adaptor, 107 is a perceptual weighting filter, 108 is a least mean square error calculator, 109 is a backward synthesis filter adaptive device, and 110 is a backward gain adaptive device. 20
Reference numeral 0 is a voice detector that inputs the parameters output from the voice encoder 100 and outputs the determination result.

In the above, the input signal h is the uniform PCM.
After the converter 101 converts the μ-law PCM into a uniform PCM, the vector buffer 102 divides the i-law PCM into a block j of five continuous input signal samples. For each input block j, the encoder has 1024 codebook vector candidates k held in the excitation codebook 103.
Is passed through the gain adjustment unit 104 and the synthesis filter 105, and from the resulting 1024 quantized signal vector candidates m, the perceptual weighting filter 107 performs frequency weighting on the input signal vector j, and weighting is performed. The root mean square error 0 is calculated as the least mean square error calculator 10
Enter 8 to determine the minimum. The 10-bit codebook index p corresponding to this optimum quantized signal vector is transmitted to the decoder.

The optimum code vector q of the least mean square error calculator 108 is then passed through the gain adjustment unit 104 and the synthesis filter 105 in order to update the filter memory in preparation for coding the next signal vector. The coefficient r of the synthesizing filter 105 and the gain s of the gain adjusting unit 104 are the back of the backward synthesizing filter adaptor 109 and the backward gain adaptor 110, respectively, based on the past quantized signal m and the gain adjusted excitation signal l. It is updated periodically by word adaptation. The period of these adaptations is such that the excitation gain s is every 1 vector (0.625 ms) and the synthesis filter coefficient r is every 4 vectors (2.5 ms). The coefficient n of the perceptual weighting filter 107 is periodically updated by the perceptual weighting filter adaptive unit 106,
The cycle is also every 4 vectors (2.5 ms), but the size of the vector buffer 102 is 1 vector (5 ms).
Therefore, the one-way delay of 2 ms or less can be realized.

Backward gain adaptor 1 of this encoder
The excitation gain s output from 10 can be considered as the input signal power, and the voice can be detected by the power determination based on this parameter. Further, the reflection coefficient t calculated inside the auditory weighting filter adaptor 106 is a vocal tract parameter indicating spectral envelope information of the input signal, and by using this, it is possible to distinguish between voice and background noise.

Therefore, the excitation gain s and the reflection coefficient t are input to the voice detector (200), and the voice detection signal u is output by comparing these two types of parameters with the prepared threshold value. You can

Next, the internal structure of the voice detector 200 is shown in FIG. 2 and will be described. Here, description will be made on the assumption that the voice detection result is output in units of 10 ms frames. The reflection coefficient t calculated by the speech encoder is the average value calculator 201.
The average value w is calculated every 10 ms (4 adaptive cycles). In the minimum error calculator 203, this average value w
With all kinds of background noise as a training sequence and all reflection coefficient vectors v output from a background noise codebook 202 with a codebook size of 8, for example, created by the LBG algorithm (known technique), for example. Then, the minimum error Eminx is output.

On the other hand, the excitation gain s calculated by the speech encoder is input to the average value calculator 204, and every 10 ms (1
The average value y of 6 adaptation periods) is calculated. The excitation gain average value y and the minimum error Eminx are input to the determination circuit 205, the presence / absence determination is performed, and the determination result z is output. In the hangover processor 206, for example, 100 ms is added to the determination result z in order to prevent word endings.
Hangover processing is performed and the final voice detection result u is output.

The decision logic flow by the decision circuit 205 will be described with reference to FIG. The excitation gain s calculated by the speech encoder is the average value y calculated by the average value calculator 204.
(Hereinafter, referred to as average excitation gain s)
The value is compared with the excitation gain threshold Ths set in. s <Ths
If is satisfied, the silence is determined in step 304, and if not satisfied, the process proceeds to the next step 302. In step 302, the minimum error Eminx calculated from the reflection coefficient t from the speech encoder is compared with the set minimum error threshold Thx,
If x <Thx is satisfied, it is determined that there is no sound in step 304, and if not satisfied, it is determined that there is sound in step 303.

Next, an example in which the present invention is applied to an actual voice signal input will be shown. FIG. 4 shows an embodiment when the S / N of the input signal h is infinite. Along with the voice signal h, the minimum error Eminx, the excitation gain s, and the final voice detection result u
Shows the change of.

When the background noise contained in the input signal is at a level that cannot be perceived as described above, voice detection using only the excitation gain s is sufficient. At this time, the excitation gain s shows a value of around 1.0 when there is no sound and about 5 or more when there is sound. If there is no background noise included in this input signal,
Naturally, there is no signal close to the nature of the signal stored in the background noise codebook 202, and the minimum error Eminx is all 0.
A value larger than 1 is shown.

FIG. 5 shows a case where a voice signal in which background noise is added to the input signal h and the S / N is about 5 dB is used. When the level of the background noise included in the input signal is relatively high as described above, voice detection using only the excitation gain s is insufficient, and in fact, the excitation gain s
Always takes a value of 20 or more. However, in this case, since the property of the background noise is similar to the property of the signal stored in the background noise codebook 202 in advance,
In the part where there is no sound, the minimum error Eminx shows a value of 0.1 or less.

In both the embodiment of FIGS. 4 and 5, the threshold values described in FIG. 3 are set as Ths = 5 and Thx = 0.1, but the voice detection by this operation operates almost accurately. I understand that.

[0046]

As described in detail above, according to the present invention, the reflection coefficient t and the excitation gain s calculated by the speech coder are used to calculate the average value within this fixed section, The presence / absence of background noise is taken into consideration because it is compared with a threshold value prepared in advance to determine the presence / absence of sound, and accurate determination is possible even in a normal use environment. Furthermore, the voice detection determination interval can be varied in 2.5 ms to 2.5 ms steps by changing the average number of two average value calculators for the reflection coefficient t and the excitation gain s. It can handle voice detection in 10 ms frame units or less. Further, since the calculation parameters of the speech coder are used, the circuit scale and the amount of calculation for speech detection are small, and there are many advantages such as suitability for LSI.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of an LD-CELP speech encoder having a speech detector of the present invention.

FIG. 2 is a configuration diagram of an embodiment of a voice detector of the present invention.

FIG. 3 is a decision logic flow of the voice detector of the present invention.

FIG. 4 is an example of a voice detection processing result of the present invention.

FIG. 5 is an example of a voice detection processing result of the present invention.

FIG. 6 is a configuration diagram of a conventional voice detector.

FIG. 7 is a decision logic flow of a conventional voice detector.

[Explanation of symbols]

100 ... LD-CELP speech encoder, 101 ... Uniform P
CM converter, 102 ... Vector buffer, 103 ... Excitation codebook, 104 ... Gain adjustment unit, 105 ... Synthesis filter, 106 ... Auditory weighting filter adaptor, 1
07 ... Auditory weighting filter, 108 ... Least mean square error calculator, 109 ... Backward synthesis filter adaptor,
110 ... Backward gain adaptor, 200 ... Speech detector, 201, 204 ... Average value calculator, 202 ... Background noise codebook, 203 ... Minimum error calculator, 205 ... Judgment circuit, 206 ... Hangover processor.

Claims (2)

[Claims] 1. A speech encoder comprising a perceptual weighting filter adaptor and a backward gain adaptor, wherein a reflection coefficient calculated by the perceptual weighting filter adaptor and an excitation calculated by the backward gain adaptor. A voice detector characterized by performing voiced / non-voiced determination by inputting each gain and comparing it with a prepared threshold value. 2. A speech coder having a speech detector, comprising the speech detector according to claim 1.