CN100419854C - Voice gain factor estimating device and method - Google Patents

Abstract

The present invention discloses an estimation device for voice gain factors, which comprises a prior signal-to-noise ratio estimation unit, a pure voice gain calculation unit, a conditional voice existence probability estimation unit and a gain factor calculation unit. The present invention simultaneously discloses an estimation method for voice gain factors. Firstly, the prior signal-to-noise ratio of input signals is calculated; then, a pure voice gain value is calculated according to the prior signal-to-noise ratio of the input signals; subsequently, the probability that voice exists in the input signals is calculated; finally, voice gain factors of the input signals are obtained through calculation combined with self-preserved noise signal reducing factors. The method estimates the voice gain factors which do not comprise integral calculation or power calculation so that complicated non-linear calculation is avoided, and the requirements for memory space and calculation quantity are reduced. The device is suitable for operation in an integer DSP chip and reduces processing delay; if the device is applied to speech, the device can satisfy the real-time requirement.

Description

A kind of voice gain factor estimating device and method

Technical field

The present invention relates to the speech enhancement technique field, be specifically related to voice gain factor estimating device and method in a kind of voice enhancing.

Background technology

Conversation often is subjected to stronger ground unrest and disturbs in the speech communication of portable terminal, and causes voice quality to descend, when serious even make the connotation that both call sides can not correct understanding the other side be expressed.At present, before communication module sent in voice, it is carried out noise reduction process usually, reach the purpose that improves voice quality by sound enhancement method.

The voice enhancement algorithm that is applied in the mobile terminal communication must have following characteristic: postpone for a short time, can not cause interference to communication; Adapt to the ground unrest that constantly changes, thereby suppress ground unrest effectively; Avoid introducing noise that artificial treatment brings as music noise etc., and guarantee that voice quality is not damaged.In addition, must be suitable in digital signal processing (DSP) chip or other dedicated processes chip, moving.

Fig. 1 realizes the theory diagram that voice strengthen in the prior art, as shown in Figure 1, at first add the Noisy Speech Signal of analyzing 11 pairs of inputs of window unit and add analysis window and handle, occur because the situation of high fdrequency component is introduced in the piecemeal processing when avoiding carrying out the FFT computing; Then, 12 pairs of Noisy Speech Signals after adding the analysis window processing of FFT arithmetic element carry out the FFT computing, obtain the coefficient Y of each Frequency point on the Noisy Speech Signal frequency domain, and each FFT coefficient Y is outputed to squared magnitude arithmetic element 13 and multiplication unit 16; Then, squared magnitude arithmetic element 13 is calculated the amplitude of each FFT coefficient Y | Y| ², and with this amplitude | Y| ²Output to noise spectrum estimation unit 14 and gain factor estimation unit 15; Then, noise spectrum estimation unit 14 is according to FFT coefficient amplitude | Y| ²The energy λ of calculating noise signal on each Frequency point _d, and with λ _dOffer gain factor estimation unit 15, afterwards, gain factor estimation unit 15 is according to λ _dWith | Y| ²Estimate voice gain factor G, and voice gain factor G is outputed to multiplication unit 16; Multiplication unit 16 multiplies each other Y and G, obtain the input signal X after noise reduction process, and X outputed to IFFT arithmetic element 17,17 couples of X of IFFT arithmetic element carry out the IFFT computing, each the IFFT coefficient that obtains outputed to add synthetic window unit 18, add to synthesize and output to overlap-add unit 19 after 18 pairs of each IFFT coefficients of window unit add synthetic window processing, the voice signal after being enhanced after the IFFT coefficient that last overlap-add unit 19 obtains each IFFT coefficient through adding synthetic window processing and the last time carries out the overlap-add computing.

In above processing procedure, adding analysis window processing, FFT computing, IFFT computing, adding that synthetic window is handled, overlap-add is handled all is technology commonly used in the digital signal processing, is not described in detail at this.And gain factor estimation unit 15 as shown in Figure 2, mainly comprises: priori voice disappearance probability estimate unit 151, priori SNR estimation unit 152, posteriority SNR estimation unit 153, condition voice exist probability estimate unit 154 and gain factor computing unit 155.The process of estimating voice gain factor is as follows:

This distributes if the FFT coefficient of noise signal d and clean speech signal x all is panel height, and the variance of its each FFT coefficient amplitude is respectively λ _d[i] and λ _x[i].For the FFT coefficient Y[i of Noisy Speech Signal on Frequency point i], establish H ₀[i] represents Y[i] on do not have voice, H ₁[i] represents Y[i] on have voice, then:

At Y[i] on when not having voice, Y[i] probability that occurs is:

$p\left(Y\right[i\left]\right|H_0\left[i\right])=\frac{1}{{\mathrm{\π\λ}}_d\left[i\right]}\exp \{-\frac{{\left|Y\right[i\left]\right|}^2}{{\mathrm{\λ}}_d\left[i\right]}\};$

At Y[i] on when having voice, Y[i] probability that occurs is:

$p\left(Y\right[i\left]\right|H_1\left[i\right])=\frac{1}{\mathrm{\π}\left({\mathrm{\λ}}_d\right[i]+{\mathrm{\λ}}_x[i\left]\right)}\exp \{-\frac{{\left|Y\right[i\left]\right|}^2}{{\mathrm{\λ}}_d\left[i\right]+{\mathrm{\λ}}_x\left[i\right]}\}---\left(1\right)$

According to bayesian criterion, as can be known, at Y[i] on exist the probability of voice to be:

$p\left(H_i\right[i\left]\right|Y\left[i\right])=\frac{\mathrm{\Λ}\left[i\right]}{1+\mathrm{\Λ}\left[i\right]}---\left(2\right)$

Wherein, $\mathrm{\Λ}\left[i\right]=\frac{1-q\left[i\right]}{q\left[i\right]}*\frac{p\left(Y\right[i\left]\right|H_1\left[i\right])}{p\left(Y\right[i\left]\right|H_0\left[i\right])},$ Q[i] at Y[i] on do not have the prior probability of voice, can obtain by the disappearance of the priori voice among Fig. 2 probability estimate unit 151.

Can obtain by formula (1) and (2) at Y[i] on exist the probability of voice to be:

$p\left[i\right]=p\left(H_1\right[i\left]\right|Y\left[i\right])={\{1+\frac{q\left[i\right]}{1-q\left[i\right]}(1+\mathrm{\ξ}[i\left]\right)*\exp (-v[i\left]\right)\}}^{-1}---\left(3\right)$

Wherein, ξ [i] is the priori signal to noise ratio (S/N ratio) of Noisy Speech Signal, can be obtained by the priori SNR estimation unit 152 among Fig. 2; $v\left[i\right]=\frac{\mathrm{\γ}\left[i\right]*\mathrm{\ξ}\left[i\right]}{1+\mathrm{\ξ}\left[i\right]},$ Wherein, γ [i] is the posteriority signal to noise ratio (S/N ratio) of Noisy Speech Signal, by posteriority SNR estimation unit 153 bases among Fig. 1 $\mathrm{\γ}\left[i\right]=\frac{{\left|Y\right[i\left]\right|}^2}{{\mathrm{\λ}}_d\left[i\right]}$ Obtain, and | Y[i] | ²By 13 outputs of squared magnitude arithmetic element, λ _d[i] is by 14 outputs of noise spectrum estimation unit.

As can be seen, Y[i] on exist the Probability p [i] of voice can have probability estimate unit 154 by the condition voice among Fig. 2 according to the ξ [i] of priori SNR estimation unit 152 outputs, the γ [i] of posteriority SNR estimation unit 153 outputs and the q[i that priori voice disappearance probability estimate unit 151 is exported] calculate.

At last, obtain through derivation:

$G_{H_1}\left[i\right]=\frac{\mathrm{\ξ}\left[i\right]}{1+\mathrm{\ξ}\left[i\right]}\exp \left(\frac{1}{2}{\∫}_{v\left[i\right]}^{\∞}\frac{e^{-t}}{t}\mathrm{dt}\right),G\left[i\right]=G_{H_1}{\left[i\right]}^{p\left[i\right]}{G_{\min }}^{1-p\left[i\right]}---\left(4\right)$

Wherein, G _MinFor noise signal is cut down the factor.

That is, gain factor computing unit 155 can obtain ν [i] according to the ξ [i] of priori SNR estimation unit 152 outputs, the γ [i] of posteriority SNR estimation unit 153 outputs, obtains the clean speech yield value according to ξ [i] and ν [i] after integral operation then Last basis

The G that self preserves _Min, there are the p[i of probability estimate unit 154 output in the condition voice] computing obtains voice gain factor G[i through exponentiation].In addition, gain factor estimation unit 155 also can with

Output to priori SNR estimation unit 152, to be used to estimate the priori signal to noise ratio (S/N ratio) ξ [i+1] on the next Frequency point (i+1).

As can be seen, in formula (4), the computing of once quadraturing, twice exponentiation computing, these two kinds of computings are because its calculated amount and memory requirements are all very big, therefore in the integer dsp chip, be very difficult to realize,, also be difficult to reach the real time execution requirement even can realize.

Summary of the invention

In view of this, fundamental purpose of the present invention is to provide a kind of voice gain factor estimating device, with the calculated amount of reduction voice gain factor estimation, and reduces the demand of voice gain factor estimation to internal memory;

Another fundamental purpose of the present invention is to provide a kind of voice gain factor method of estimation, to reduce the computational complexity that voice gain factor is estimated.

For achieving the above object, technical scheme of the present invention is achieved in that

A kind of voice gain factor estimating device, this device comprises:

Priori SNR estimation unit is used to estimate the priori signal to noise ratio (S/N ratio) of input signal, and will estimates that the priori signal to noise ratio (S/N ratio) that obtains outputs to clean speech gain calculating unit;

Clean speech gain calculating unit is used for the priori signal to noise ratio (S/N ratio) according to the output of priori SNR estimation unit, calculates the clean speech yield value, and this clean speech yield value is outputed to the gain factor computing unit;

There is the probability estimate unit in the condition voice, are used to calculate the probability that has voice on the input signal, and will exist the probability of voice to output to the gain factor computing unit on this input signal;

The gain factor computing unit is used for cutting down the factor, computing voice gain factor according to probability that has voice on clean speech yield value, the input signal and noise signal.

Described clean speech gain calculating unit comprises: add division operation unit and extraction of square root arithmetic element, wherein,

Add the division operation unit, be used for the priori signal to noise ratio (S/N ratio) of priori SNR estimation unit output divided by this priori signal to noise ratio (S/N ratio) and 1 with, the result who obtains is outputed to the extraction of square root arithmetic element;

The extraction of square root arithmetic element is used for that the computing of extracting square root obtains the clean speech yield value to the value that adds division operation unit output, and this clean speech yield value is outputed to the gain factor computing unit.

Described gain factor computing unit comprises: multiplication unit, noise-cut factor output unit, subtracts the multiplication unit and adds arithmetic element, wherein,

The multiplication unit is used for adding arithmetic element with having the probability that there is voice on the input signal of probability estimate unit from the condition voice and multiply each other from the clean speech yield value of clean speech gain calculating unit, the value that obtains being outputed to;

Noise-cut factor output unit, the noise signal reduction factor that is used for self preserving outputs to and subtracts the multiplication unit;

Subtract the multiplication unit, be used for existing the probability of voice to subtract each other with existing on the input signal of probability estimate unit from the condition voice with 1, and the noise signal of the difference that will obtain and noise-cut factor output unit output cuts down the factor and multiplies each other, and the value that obtains outputed to add arithmetic element;

Add arithmetic element, be used for the value of multiplication unit output is carried out addition with the value that subtracts the output of multiplication unit, obtain voice gain factor.

Described clean speech gain calculating unit is further used for, and the clean speech yield value is outputed to priori SNR estimation unit.

A kind of voice gain factor method of estimation, this method comprises:

A, calculate the priori signal to noise ratio (S/N ratio) of input signal, and according to the clean speech yield value of this priori snr computation input signal;

The probability that has voice on B, the calculating input signal;

C, cut down the factor, calculate the voice gain factor of input signal according to probability that has voice on the clean speech yield value of input signal, the input signal and noise signal.

The clean speech yield value of the described calculating input signal of steps A is specially:

${\mathrm{<figure-callout\; id="1"\; label="G\; H"\; filenames="C20051011506000151.png"\; state="\{\{state\}\}">G</figure-callout>}}_{H_{}}={\left(\frac{\mathrm{\&xi;}}{\mathrm{\&xi;}+1}\right)}^{\frac{1}{2}},$ Wherein,

Be the clean speech yield value of described input signal, ξ is the priori signal to noise ratio (S/N ratio) of described input signal.

Described step B is specially:

Calculate the probability q that there are not voice in input signal, and calculate the posteriority signal to noise ratio (S/N ratio) γ of input signal, then exist the Probability p of voice to be on the input signal:

${p=(1-\frac{q}{1-q}(1+\mathrm{\&xi;})*\exp (-v))}^{-1},$ Wherein, $v=\frac{\mathrm{\&gamma;}*\mathrm{\&xi;}}{1+\mathrm{\&xi;}},$ ξ is the priori signal to noise ratio (S/N ratio) of described input signal.

Described step C is specially:

${G=p*\left(\frac{\mathrm{\&xi;}}{\mathrm{\&xi;}+1}\right)}^{\frac{1}{2}}+(1-p)*G_{\min },$ Wherein, G is the voice gain factor of described input signal, and p is the probability that has voice on the described input signal of step B, and ξ is the priori signal to noise ratio (S/N ratio) of the described input signal of steps A, G _MinFor described noise signal is cut down the factor.

Compared with prior art, voice gain factor estimating device provided by the present invention and method are passed through the priori signal to noise ratio (S/N ratio) according to the input signal of priori SNR estimation unit output, calculate the clean speech yield value of input signal, there is the probability that has voice on the input signal of probability estimate unit output according to this clean speech yield value and condition voice then, and noise signal is cut down the factor, calculate the voice gain factor of input signal, realized estimation to voice gain factor, and voice gain factor does not contain integration and exponentiation computing, avoided complicated nonlinear operation, reduced requirement to memory space and calculated amount, be suitable in the integer dsp chip, moving, and reduced processing delay, but be applied in requirement of real time in the voice enhancing.

Description of drawings

Fig. 1 realizes the theory diagram that voice strengthen in the prior art;

Fig. 2 is the voice gain factor estimation principles block diagram in the existing voice enhancing;

Fig. 3 is a process flow diagram of estimating voice gain factor during voice provided by the invention strengthen;

Fig. 4 is the device block diagram one that the voice gain factor during voice provided by the invention strengthen is estimated;

Fig. 5 is the device block diagram two that the voice gain factor during voice provided by the invention strengthen is estimated.

Embodiment

When there are voice in input signal Y [i] on Frequency point i,, then have: Y because voice signal is uncorrelated with noise signal ²[i]=X ²[i]+N ²[i], wherein, X[i] be the voice signal on Frequency point i, can the FFT coefficient etc. expression; N[i] be the noise signal on Frequency point i, can the FFT coefficient etc. expression; Y[i] be the input signal on Frequency point i, can the FFT coefficient etc. expression, then:

$\frac{X^2\left[i\right]}{Y^2\left[i\right]}=\frac{X^2\left[i\right]}{X^2\left[i\right]+N^2\left[i\right]}\stackrel{A}{=}\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1},$ Wherein, ξ [i] is the priori signal to noise ratio (S/N ratio) of input signal on Frequency point i.

Then have: ${X\left[i\right]=\left(\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1}\right)}^{\frac{1}{2}}*Y\left[i\right],$ Be that the clean speech yield value is: $G_{H_1}\left[i\right]={\left(\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1}\right)}^{\frac{1}{2}},$

Combined input signal Y[i] there is the Probability p [i] of voice, can obtain voice gain factor ${G\left[i\right]:G\left[i\right]=p\left[i\right]*\left(\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1}\right)}^{\frac{1}{2}}+(1-p[i\left]\right)*G_{\min },$ Wherein, G _MinFor noise signal is cut down the factor, and 0＜G _Min＜1.

The estimated value of the voice signal on the Frequency point i then, the voice signal after promptly strengthening

For:

$\hat{X}\left[i\right]=\left(p\right[i]*G_{H_1}[i]+(1-p[i\left]\right)*G_{\min })*Y\left[i\right].$

From the above mentioned, carry out flow process that voice gain factor estimates during voice provided by the invention strengthen as shown in Figure 3, its concrete steps are as follows:

Step 301: calculate the priori signal to noise ratio (S/N ratio) ξ [i] of input signal on each Frequency point i.

Here, ξ [i] can adopt prior art to calculate.

Step 302:, calculate the clean speech yield value of input signal on each Frequency point i according to priori signal to noise ratio (S/N ratio) ξ [i] $G_{H_1}\left[i\right]:G_{H_1}\left[i\right]={\left(\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1}\right)}^{\frac{1}{2}}.$

Step 303: calculate there are voice in input signal on each Frequency point i Probability p [i].

Here, p[i] can adopt prior art to calculate, as: ${p\left(i\right)=\{1+\frac{q\left[i\right]}{1-q\left[i\right]}(1-\mathrm{\&xi;}[i\left]\right)*\exp (-v[i\left]\right)\}}^{-1},$ Wherein, q[i] be there are not voice in input signal on Frequency point i probability, $v\left[i\right]=\frac{\mathrm{\&gamma;}\left[i\right]*\mathrm{\&xi;}\left[i\right]}{1+\mathrm{\&xi;}\left[i\right]},$ ξ [i] is that priori signal to noise ratio (S/N ratio), γ [i] are the posteriority signal to noise ratio (S/N ratio).

Step 304: according to

, p[i] and noise signal cut down factor G _Min, calculate the voice gain factor of input signal on each Frequency point i ${G\left[i\right]:G\left[i\right]=p\left[i\right]*\left(\frac{\mathrm{\&xi;}\left[i\right]}{\mathrm{\&xi;}\left[i\right]+1}\right)}^{\frac{1}{2}}+(1-p[i\left]\right)*G_{\min }.$

Afterwards, according to $\hat{X}\left[i\right]=Y\left[i\right]*G\left[i\right],$ Can obtain the estimated value of the voice signal on Frequency point i Promptly obtain the voice signal after the enhancing on the Frequency point i,

Can the FFT coefficient etc. expression.After this, right By frequency-time-domain-transformation, time domain voice strengthen as: add after synthetic window and overlap-add etc. handle, just can obtain the voice signal after the enhancing on the time domain.

In the practical application, the extraction of square root computing can approach as realizations such as newton's iteration, direct current approximatiosses by fast zoom table, segmented line shape function, can realize in the integer dsp chip.

Fig. 4 is the device block diagram one that the voice gain factor during realization voice provided by the invention strengthen is estimated, as shown in Figure 4, it mainly comprises:

Priori SNR estimation unit 41: be used to estimate the priori signal to noise ratio (S/N ratio) ξ [i] of input signal on Frequency point i, and ξ [i] is outputed to clean speech gain calculating unit 42.

Priori SNR estimation unit 41 can adopt prior art to obtain ξ [i].

Clean speech gain calculating unit 42: be used for ξ [i], calculate the clean speech yield value according to 41 outputs of priori SNR estimation unit

And will

Output to gain factor computing unit 44.

Further, clean speech gain calculating unit 42 can with

Output to priori SNR estimation unit 41, to be used to estimate the priori signal to noise ratio (S/N ratio) ξ [i+1] of input signal on Frequency point (i+1).

There is probability estimate unit 43 in the condition voice: be used to calculate there are voice in input signal on Frequency point i Probability p [i], and with p[i] output to gain factor computing unit 44.

The condition voice exist probability estimate unit 43 can adopt prior art to obtain p[i].

Gain factor computing unit 44: be used for exporting according to clean speech gain calculating unit 42

There is the p[i of probability estimate unit 43 output in the condition voice] and the noise signal reduction factor G that self preserves _Min, computing voice gain factor G[i].

In actual applications, voice gain factor estimating device provided by the invention is obtaining voice gain factor G[i] after, can be with as shown in Figure 1 IFFT unit 17, the unit that is used to realize as shown in Figure 1 multiplication unit 16, voice strengthen on the frequency domain unit, is used to realize frequency domain-time domain conversion, be used to realize that add synthetic window unit 18 and the overlap-add unit 19 etc. as shown in Figure 1, unit that the voice on the time domain strengthen finish the voice enhancement process.

Further, as shown in Figure 5, clean speech gain calculating unit 42 comprises:

Add division operation unit 421: be used for the priori signal to noise ratio (S/N ratio) ξ [i] of priori SNR estimation unit 41 outputs is made the following division operation that adds:

And the value that will obtain outputs to extraction of square root arithmetic element 422.

Extraction of square root arithmetic element 422: be used for the value that the adds division operation unit 421 output computing of extracting square root is obtained the clean speech yield value

And with this clean speech yield value

Output to gain factor computing unit 44.

Further, as shown in Figure 5, gain factor computing unit 44 comprises:

Multiplication unit 441: the input signal that is used for having probability estimate unit 43 from the condition voice exists the Probability p [i] of voice and the clean speech yield value of clean speech gain calculating unit 42 outputs on Frequency point i Multiply each other promptly: calculate

The value that obtains outputed to add arithmetic element 444.

Noise-cut factor output unit 442: the noise signal that is used for self preserving is cut down factor G _MinOutput to and subtract multiplication unit 443.

Subtract multiplication unit 443: be used on Frequency point i, existing the Probability p [i] of voice and the noise signal of noise-cut factor output unit 442 outputs to cut down factor G to the input signal that has probability estimate unit 43 from the condition voice _MinSubtract multiplication promptly: calculate (1-p[i]) * G _Min, and the value that will obtain outputs to and adds arithmetic element 444.

Add arithmetic element 444: be used for the value of multiplication unit 441 outputs is carried out addition with the value that subtracts 443 outputs of multiplication unit, obtain voice gain factor G[i].

In practice, the present invention is applied in the voice enhancing, the processing delay that voice strengthen can be controlled within 32 milliseconds.

The above only is process of the present invention and method embodiment, in order to restriction the present invention, all any modifications of being made within the spirit and principles in the present invention, is not equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1. a voice gain factor estimating device is characterized in that, this device comprises:

Priori SNR estimation unit is used to estimate the priori signal to noise ratio (S/N ratio) of input signal, and will estimates that the priori signal to noise ratio (S/N ratio) that obtains outputs to clean speech gain calculating unit;

Clean speech gain calculating unit is used for the priori signal to noise ratio (S/N ratio) according to the output of priori SNR estimation unit, calculates the clean speech yield value, and this clean speech yield value is outputed to the gain factor computing unit;

There is the probability estimate unit in the condition voice, are used to calculate the probability that has voice on the input signal, and will exist the probability of voice to output to the gain factor computing unit on this input signal;

The gain factor computing unit is used for cutting down the factor, computing voice gain factor according to probability that has voice on clean speech yield value, the input signal and noise signal.

2. device as claimed in claim 1 is characterized in that, described clean speech gain calculating unit comprises: add division operation unit and extraction of square root arithmetic element, wherein,

Add the division operation unit, be used for the priori signal to noise ratio (S/N ratio) of priori SNR estimation unit output divided by this priori signal to noise ratio (S/N ratio) and 1 with, the result who obtains is outputed to the extraction of square root arithmetic element;

The extraction of square root arithmetic element is used for that the computing of extracting square root obtains the clean speech yield value to the value that adds division operation unit output, and this clean speech yield value is outputed to the gain factor computing unit.

3. device as claimed in claim 1 or 2 is characterized in that, described gain factor computing unit comprises: multiplication unit, noise-cut factor output unit, subtract the multiplication unit and add arithmetic element, wherein,

The multiplication unit is used for adding arithmetic element with having the probability that there is voice on the input signal of probability estimate unit from the condition voice and multiply each other from the clean speech yield value of clean speech gain calculating unit, the value that obtains being outputed to;

Noise-cut factor output unit, the noise signal reduction factor that is used for self preserving outputs to and subtracts the multiplication unit;

Subtract the multiplication unit, be used for existing the probability of voice to subtract each other with existing on the input signal of probability estimate unit from the condition voice with 1, and the noise signal of the difference that will obtain and noise-cut factor output unit output cuts down the factor and multiplies each other, and the value that obtains outputed to add arithmetic element;

Add arithmetic element, be used for the value of multiplication unit output is carried out addition with the value that subtracts the output of multiplication unit, obtain voice gain factor.

4. device as claimed in claim 1 is characterized in that, described clean speech gain calculating unit is further used for, and the clean speech yield value is outputed to priori SNR estimation unit.

5. a voice gain factor method of estimation is characterized in that, this method comprises:

A, calculate the priori signal to noise ratio (S/N ratio) of input signal, and according to the clean speech yield value of this priori snr computation input signal;

The probability that has voice on B, the calculating input signal;

C, cut down the factor, calculate the voice gain factor of input signal according to probability that has voice on the clean speech yield value of input signal, the input signal and noise signal.

6. method as claimed in claim 5 is characterized in that, the clean speech yield value of the described calculating input signal of steps A is specially:

$G_{H_i}={\left(\frac{\mathrm{\&xi;}}{\mathrm{\&xi;}+1}\right)}^{\frac{1}{2}},$ Wherein, Be the clean speech yield value of described input signal, ξ is the priori signal to noise ratio (S/N ratio) of described input signal.

7. method as claimed in claim 5 is characterized in that, described step B is specially:

Calculate the probability q that there are not voice in input signal, and calculate the posteriority signal to noise ratio (S/N ratio) γ of input signal, then exist the Probability p of voice to be on the input signal:

$p={\{1+\frac{q}{1-q}(1+\mathrm{\&xi;})*\exp (-v)\}}^{-1},$ Wherein, $v=\frac{\mathrm{\&gamma;}*\mathrm{\&xi;}}{1+\mathrm{\&xi;}},$ ξ is the priori signal to noise ratio (S/N ratio) of described input signal.

8. method as claimed in claim 5 is characterized in that, described step C is specially:

$G=p*{\left(\frac{\mathrm{\&xi;}}{\mathrm{\&xi;}+1}\right)}^{\frac{1}{2}}+(1-p)*G_{\min },$ Wherein, G is the voice gain factor of described input signal, and p is the probability that has voice on the described input signal of step B, and ξ is the priori signal to noise ratio (S/N ratio) of the described input signal of steps A, G _MinFor described noise signal is cut down the factor.

Images