CN105679330A - Digital hearing aid noise reduction method based on improved sub-band signal-to-noise ratio estimation - Google Patents

Abstract

The invention provides a digital hearing aid noise reduction method based on improved sub-band signal-to-noise ratio estimation, the method comprises the steps: a decomposition filter decomposes an original signal into a plurality of sub-bands; a cross-correlation function and each mean square value of two adjacent frame signals in each sub-band is calculated, so that a signal-to-noise ratio of the sub-band is estimated; gain of each sub-band is calculated according to the estimated signal-to-noise ratio, and the gain multiplied by a sub-band signal is a corrected sub-band signal; finally, the corrected sub-band signals are combined to obtain a noise reduction processed voice. According to the invention, the accurater sub-band signal-to-noise ratio estimation method makes the background noise inhibitory effect being better, and auditory fatigue of hearing-aid users is reduced; the method is simple and has high efficiency, inverse Fourier transform is avoided, so that the time delay performance is greatly improved, the method improved for 60.6% from a traditional spectral subtraction, and the method improved for 40.7% from a traditional Wiener filtering method.

Description

Based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation

Technical field

The present invention relates to a kind of based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation.

Background technology

Hearing loss is the serious disease affecting human lives. Long-term dysaudia, except affecting daily exchange, also can cause psychological problems, brings white elephant to society and family. Wear hearing aid is that light-severe listens the maximally effective audition intervention of barrier patient and the means of rehabilitation. Noise circumstance is understood the most FAQs that voice difficulty is puzzlement sonifer user. In order to improve speech understanding degree and listen to comfort level, sonifer speech enhancement technique substantially can be attributed to two classes: shotgun microphone and noise-reduction method. The former is based on voice and noise diversity design spatially, and utilization orientation mike or beam-forming technology carry out the voice signal on Enhancement feature direction. But this type of method is subject to quantity or the size limitation of mike, and performance improvement is limited, and is not suitable for deep auditory meatus sonifer. Difference on time that second method is intended to utilize voice and noise and frequency spectrum, separates voice from signals and associated noises. But, voice and noise are likely on time and frequency spectrum there is overlap, and therefore a lot of research worker conduct in-depth research for this problem.

In speech enhan-cement, researcher proposes the significant method of a lot of noise reduction effect. Patent US2012/8204263B2 " Methodofestimatingweightingfunctionofaudiosignalsinahear ingaid " i.e. is a kind of sonifer audio signal Enhancement Method, utilize multi-microphone Collect jointly ambient sound, realize speech enhan-cement effect, but for wearing in auditory meatus or the patient of neighbouring small-sized custom hearing aid, it is but inapplicable that this class carries out speech enhan-cement by multi-microphone; Patent US2014/13198739.8, " Enhanceddynamicsprocessingofstreamingaudiobysourcesepara tionandremixing " i.e. be a kind of enhancing audio stream method using source to separate and mix, employing Non-negative Matrix Factorization is theoretical, by clean speech data and noise data being carried out study voice and noise model, but the method complexity and the requirement of equipment is higher; Patent CN200510086877.8 " sound enhancement method for sonifer " utilizes auditory masking curve that filtering parameter is adjusted, but the method needs further appreciating that human auditory system physiological models understanding, and the method usually contains the computing of some complexity such as substantial amounts of exponent arithmetic, logarithm operation.Although these above-mentioned methods have certain speech enhan-cement effect, but it is not particularly suited in the significantly high sonifer of real-time and power consumption requirements.

Summary of the invention

It is an object of the invention to provide a kind of based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, solve that sonifer speech enhancement technique in prior art exists or by mike quantity or size limitation, performance improvement is limited, and be not suitable for deep auditory meatus sonifer, or the problem that voice and noise are likely to there is overlap on time and frequency spectrum.

In order to realize above-mentioned target, the present invention adopts the following technical scheme that:

A kind of based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, including:

Adopt analysis filter bank that original signal resolves into several subbands, then calculate the cross-correlation function of adjacent two frame signals in each subband and respective mean-square value, estimate the signal to noise ratio of this subband;

Secondly, according to each subband gain of the signal-to-noise ratio computation estimated, and the subband signal obtaining revising that is multiplied with subband signal;

Finally, revised each subband signal is comprehensively obtained the voice after noise reduction process.

Further, following steps are specifically included:

S1, microphone input signal x (n) to digital deaf-aid are AD converted, and by the digital signal framing x after conversion_i(n) (i=0,1,2 ... N-1), wherein N is frame length;

S2, pass through resolution filter: by every frame signal by multichannel 6 rank IIR analysis filter bank h_i(n) (i=0,1 ... M-1), M is port number, and passband ripple 0.5dB, with c_iCarry out down-sampled for down-sampled coefficient, obtain M subband signal: x (i, m) (i=0,1,2 ... N-1) (m=0,1,2 ... M-1);

S3, subband signal-to-noise ratio (SNR) estimation: the signal-to-noise ratio (SNR) estimation value of m-th subband the i-th frame is

$SNR(i,m)=10log\left(\frac{\mathrm{\α}P(i,m)+(1-\mathrm{\α})P(i-1,m)}{{\mathrm{\σ}}^2\left(m\right)}\right)---\left(12\right)$

In formula, (i, m), (i-1, m) for the pure voice signal power of the in m-th subband i-th and i-th-1 frame, σ for P for P²M () is the noise variance in m-th subband, α is smoothing factor, 0 < α < 1.

S4, yield value calculate: by the signal to noise ratio snr that estimates, (i m) brings gain function g into_dB(i, in m), and by g_db(i, m) transform to amplitude domain g (i, m).

S5, subband noise suppress: primary speech signal is carried out noise attentuation by the yield value of each subband by obtaining, and obtains enhanced voice signal: y_im(n)=g (i, m) x (i, m);

S6, subband signal comprehensively export: by the output signal y of M passage_im(n) (i=0,1 ... M-1) first with c_iCarry out a liter sampling for liter downsampling factor, be finally synthesizing voice signal y (n) obtaining output.

Further, in step S3, subband signal-to-noise ratio (SNR) estimation particularly as follows:

The i-th and i-th-1 frame signal on S31, definition kth Frequency point is adjacent signals vector X_adjoin(i, k), calculate adjacent signals vector autocorrelation matrix R (i, k);

S32, utilize L frequency spectrum in each subband to estimate correlation matrix R (i, the E [X in m)²(i,m)]、E[X²(i-1, m)] and E [X (i-1, m) X (i, m)], E [X²(i,m)]、E[X²(i-1, m)] mean-square value of respectively the i-th frame in m-th subband and the i-th-1 frame, E [X (i, m) X (i-1, m)] represent the i-th frame and the cross-correlation function of the i-th-1 frame;

S33, solve the i-th and i-th-1 frame in adjacent two frame voice signals and m-th subband pure voice signal power P (i, m), P (i-1, m), and the noise variance σ in m-th subband²(m), for estimate m-th subband the i-th frame signal-to-noise ratio (SNR) estimation value SNR (i, m).

Further, in step S31, the autocorrelation matrix calculating adjacent signals vector is as follows:

$R(i,k)=E\&lsqb;X_{adjoin}(i,k)X_{adjoin}{(i,k)}^H\&rsqb;=\left(\begin{array}{cc}E\left(X^2\right(i,k\left)\right)& E\left(X\right(i,k\left)X\right(i-1,k\left)\right)\\ E\left(X\right(i,k\left)X\right(i-1,k\left)\right)& E\left(X^2\right(i-1,k\left)\right)\end{array}\right)---\left(1\right);$

Wherein, adjacent signals vector X_adjoin(i, k)=X (i, k), X (i-1, k) }^T, E (X²(i, k)), E (X²The i-th frame in (i-1, k)) respectively kth Frequency point kth subband and the mean-square value of the i-th-1 frame, ((i, k) X (i-1, k)) represents the i-th frame and the cross-correlation function of the i-th-1 frame to X to E.

Further, in step S32, after signal wave filter by analysis obtains M subband, according to each subband comprise L=N/M frequency spectrum estimate m-th subband consecutive frame correlation matrix R (i, m):

$R(i,m)=\left(\begin{array}{cc}E\&lsqb;X^2(i,m)\&rsqb;& E\&lsqb;X(i,m)X(i-1,m)\&rsqb;\\ E\&lsqb;X(i,m)X(i-1,m)\&rsqb;& E\&lsqb;X^2(i-1,m)\&rsqb;\end{array}\right)---\left(2\right)$

In formula (2), E [X²(i,m)]、E[X²(i-1, m)] mean-square value of respectively the i-th frame in m-th subband and the i-th-1 frame, E [X (i, m) X (i-1, m)] represent the i-th frame and the cross-correlation function of the i-th-1 frame.

Further, in step S4, gain function is

g_dB(i, m)=λ_im(SNR)f(x_dB(n))(13)

Wherein, g_db(i m) represents the gain of the i-th frame voice signal in m subband.

Further, in step S4, by the gain g of the i-th frame voice signal in m subband_db(i, m) transform to amplitude domain g (i, m):

$g(i,m)={10}^{2\&CenterDot;g_{dB}(i,m)}---\left(14\right).$

The invention has the beneficial effects as follows: this kind is based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, by the signal-to-noise ratio (SNR) estimation to subband signal, and the yield value of each subband is calculated according to subband state of signal-to-noise, achieve the signals and associated noises to quiet signal or high signal noise ratio and only carry out slight noise abatement, and the noise cancellation signal that has of low signal noise ratio is improved reduction, alleviate the auditory fatigue of user, and solve prior art and be difficult to be applicable to the difficulty of the sonifer that real-time and power consumption requirements is significantly high because computation complexity is high. This kind is based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, and subband signal-noise ratio estimation method makes background noise inhibition better more accurately, alleviates the auditory fatigue of hearing aid user; Method is simply efficient, it is to avoid inverse Fourier transform (IFFT) makes time delay performance obtain to improve largely, and more traditional spectrum-subtraction improves 60.6%, more traditional Wiener Filter Method improves 40.7%.

Accompanying drawing explanation

Fig. 1 is the embodiment of the present invention schematic flow sheet based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation;

Fig. 2 is the explanation schematic diagram of digital deaf-aid noise reduction model in embodiment;

Fig. 3 is the linear-proportional gain's function revised in embodiment;

Fig. 4 is each algorithm delay performance contrast in embodiment.

Detailed description of the invention

The preferred embodiments of the present invention are described in detail below in conjunction with accompanying drawing.

Current digital deaf-aid adopts multiple-channels algorithm mostly, and voice signal is divided into some subbands according to frequency, and each subband individually carries out noise suppressed. The wherein difference according to frequency band division methods, can be divided into wide and non-wide two kinds of methods. Because the perception of sound ability of different frequency range is different by human ear, therefore it is generally adopted non-wide frequency band division methods, the auditory properties of human ear cochlea can be simulated better. Additionally the bank of filters of threshold sampling shows poor in this performance of stopband attenuation, therefore it is better to be generally adopted noise suppressed performance, the over-sampling bank of filters of flexible design. The frequency band division methods that right and wrong that the frequency band division methods of embodiment adopts are wide, additionally bank of filters adopts over-sampling wave filter.

Embodiment

A kind of based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, such as Fig. 1, adopt 6 rank IIR analysis filter bank that original signal resolves into 16 subbands, then calculate the cross-correlation function of adjacent two frame signals in each subband and respective mean-square value, estimate the signal to noise ratio of this subband;Secondly, according to each subband gain of the signal-to-noise ratio computation estimated, and the subband signal obtaining revising that is multiplied with subband signal; Finally, revised each subband signal is comprehensively obtained the voice after noise reduction process.

Specifically include following steps:

S1, microphone input signal x (n) to digital deaf-aid are AD converted, and by the digital signal framing x after conversion_i(n) (i=0,1,2 ... N-1), wherein N is frame length;

S2, pass through resolution filter: by every frame signal by multichannel 6 rank IIR analysis filter bank h_i(n) (i=0,1 ... M-1), M is port number, and passband ripple 0.5dB, with c_iCarry out down-sampled for down-sampled coefficient, obtain M subband signal: x (i, m) (i=0,1,2 ... N-1) (m=0,1,2 ... M-1);

S3, subband signal-to-noise ratio (SNR) estimation: the signal-to-noise ratio (SNR) estimation value of m-th subband the i-th frame is

$SNR(i,m)=10log\left(\frac{\mathrm{\&alpha;}P(i,m)+(1-\mathrm{\&alpha;})P(i-1,m)}{{\mathrm{\&sigma;}}^2\left(m\right)}\right)---\left(12\right)$

In formula, (i, m), (i-1, m) for the pure voice signal power of the in m-th subband i-th and i-th-1 frame, σ for P for P²M () is the noise variance in m-th subband, α is smoothing factor, takes 0.5 in the present embodiment.

S4, yield value calculate: by the signal to noise ratio snr that estimates, (i m) brings gain function g into_dB(i, in m), and by g_db(i, m) transform to amplitude domain g (i, m).

S5, subband noise suppress: primary speech signal is carried out noise attentuation by the yield value of each subband by obtaining, and obtains enhanced voice signal: y_im(n)=g (i, m) x (i, m);

S6, subband signal comprehensively export: by the output signal y of M passage_im(n) (i=0,1 ... M-1) first with c_iCarry out a liter sampling for liter downsampling factor, be finally synthesizing voice signal y (n) obtaining output.

Embodiment based on improve subband signal-to-noise ratio (SNR) estimation digital deaf-aid noise-reduction method as in figure 2 it is shown, Fig. 2 digital deaf-aid noise reduction model in M be total number of channels, frequency band is divided into 16 subbands according to auditory perceptual characteristic and cochlea filter characteristic by embodiment:

Wherein, f_iRepresent the lower sideband frequencies of the i-th subband,Representing that the upper side band cymometer of the i-th subband calculates the lower sideband frequency of each subband, result of calculation is as shown in table 1.

Table 1 frequency band division result

In Fig. 2, x₀(n),x₁(n),…x_M-1N () is through h_iN the effect of () is that original signal is decomposed M the subband signal obtained, and keep original sampling frequency, h_iN () also has the effect of frequency overlapped-resistable filter simultaneously, embodiment adopts 3 order modes to intend Chebyshev's I type and designs 6 rank IIR resolution filter h_i(n) (i=0,1 ... 15), passband ripple 0.5dB. c_iBe the i-th subband/liter downsampling factor drops. As downsampling factor c_iDuring=M, it is called threshold sampling; c_iDuring < M, it is called over-sampling. The present embodiment uses over-sampling. u₀(n),u₁(n),…u_M-1N () is respectively through with c₀(n),c₁(n),…c_M-1N () carries out down-sampled M the subband signal obtained for downsampling factor, down-sampled coefficient is as shown in table 2. ;

The down-sampled coefficient of table 2

G in Fig. 2₀(n),g₁(n),…g_M-1N () is the amount of gain of each subband, respectively with u₀(n),u₁(n),…u_M-1N () is multiplied and obtains strengthening the output signal v after algorithm process through sub-band speech₀(n),v₁(n),…v_M-₁(n); V after process₀(n),v₁(n),…v_M-1N () signal is then through with c₀(n),c₁(n),…c_M-1N () carries out a liter sampling for downsampling factor. Each subband signal final is comprehensive, the clean speech signal of output estimation.

Subband signal-to-noise ratio (SNR) estimation

Owing to voice signal and noise are time varying signal, in each frame, the frequency domain distribution of voice and noise energy is unbalanced, it is therefore desirable to estimate the signal to noise ratio of signal in real time. The present invention proposes the signal-noise ratio estimation method of a kind of improvement, utilizes the autocorrelation matrix of adjacent two frame noisy speech signal, estimates the signal to noise ratio of each subband.

Subband signal-to-noise ratio (SNR) estimation is specific as follows:

Assume that the i-th frame signal is:

X (i, n)=s (i, n)+n (i, n) (2)

In formula, (i, n) represents Noisy Speech Signal to x, and (i, n) represents primary speech signal to s, and (i n) represents noise signal to n.

Its Fourier transformation is:

X (i, k)=S (i, k)+N (i, k) (3)

The i-th and i-th-1 frame signal on definition kth Frequency point is adjacent signals vector X_adjoin(i, k), then

X_adjoin(i, k)=X (i, k), X (i-1, k) }^T

Calculate the autocorrelation matrix of adjacent signals vector:

$R(i,k)=E\&lsqb;X_{adjoin}(i,k)X_{adjoin}{(i,k)}^H\&rsqb;=\left(\begin{array}{cc}E\left(X^2\right(i,k\left)\right)& E\left(X\right(i,k\left)X\right(i-1,k\left)\right)\\ E\left(X\right(i,k\left)X\right(i-1,k\left)\right)& E\left(X^2\right(i-1,k\left)\right)\end{array}\right)---\left(4\right)$

If only with X, (i, k) (i-1, k) estimates the correlation matrix of each frequency, can produce bigger estimation difference with X. After signal wave filter by analysis obtains M subband, it is possible to comprise L according to each subband_mIndividual frequency spectrum estimate m-th subband consecutive frame correlation matrix R (i, m):

$R(i,m)=\left(\begin{array}{cc}E\&lsqb;X^2(i,m)\&rsqb;& E\&lsqb;X(i,m)X(i-1,m)\&rsqb;\\ E\&lsqb;X(i,m)X(i-1,m)\&rsqb;& E\&lsqb;X^2(i-1,m)\&rsqb;\end{array}\right)---\left(5\right)$

In formula (6), E [X²(i,m)]、E[X²(i-1, m)] mean-square value of respectively the i-th frame in m-th subband and the i-th-1 frame, E [X (i, m) X (i-1, m)] represent the i-th frame and the cross-correlation function of the i-th-1 frame.

Assuming that clean speech signal s (t) of any frame is unrelated with the noise signal of any frame statistics, the noise of consecutive frame is uncorrelated simultaneously, then every in formula (6) can abbreviation be

E[X²(i, m)]=P (i, m)+σ²(m)(6)

E[X²(i-1, m)]=P (i-1, m)+σ²(m)(7)

Wherein, (i, m), (i-1, m) for the pure voice signal power of the in m-th subband i-th and i-th-1 frame, σ for P for P²M () is the noise variance in m-th subband.

E [X (i, m) X (i-1, m)]=E [S (and i, m) S (i-1, m)] (7+)

Above formula is deformed, can obtain

$\begin{array}{c}E\&lsqb;X(i,m)X(i-1,m)\&rsqb;=E\&lsqb;\frac{S(i,m)}{S(i-1,m)}{S}^2(i-1,m)\&rsqb;\\ =\mathrm{\&lambda;}(i-1,m)E\&lsqb;S^2(i-1,m)\&rsqb;\\ =\mathrm{\&lambda;}(i-1,m)P(i-1,m)\end{array}---\left(8\right)$

From above formula

λ (i-1, m)=S (i, m)/S (i-1, m)=(X (i, m)-N (i, m))/(X (i-1, m)-N (i-1, m))

(8+)

Assuming that noise is smooth change, (i, m) with N (i-1, m) the noise spectrum N for estimating time quiet for N_silence(m)。

(i m) may finally be expressed as such R

$R(i,m)=\left(\begin{array}{cc}P(i,m)+{\mathrm{\&sigma;}}^2\left(m\right)& \mathrm{\&lambda;}(i-1,m)P(i-1,m)\\ \mathrm{\&lambda;}(i-1,m)P(i-1,m)& P(i-1,m)+{\mathrm{\&sigma;}}^2\left(m\right)\end{array}\right)---\left(9\right)$

Utilize the L in each subband_mIndividual frequency spectrum estimates the E [X in (7) (8) (9) formula²(i, m)], E [X²(i-1, m)], E [X (and i-1, m) X (i, m)].

$E\&lsqb;X^2(i,m)\&rsqb;=\frac{1}{L_m}\underset{k=(m-1)L+1}{\overset{mL}{\&Sigma;}}{X}^2(i,k)---\left(10\right)$

$E\&lsqb;X^2(i-1,m)\&rsqb;=\frac{1}{L_m}\underset{k=(m-1)L+1}{\overset{mL}{\&Sigma;}}{X}^2(i-1,k)---\left(11\right)$

$E\&lsqb;X(i,m)X(i-1,m)\&rsqb;=\frac{1}{L_m}\underset{k=(m-1)L+1}{\overset{mL}{\&Sigma;}}X(i,k)X(i-1,k)---\left(12\right)$

Bring in formula (7), (8), (9), and P (i, m), σ²M (), (i-1 m) can uniquely determine P.

The signal-to-noise ratio (SNR) estimation value of such m-th subband the i-th frame is

$SNR(i,m)=10log\left(\frac{\mathrm{\&alpha;}P(i,m)+(1-\mathrm{\&alpha;})P(i-1,m)}{{\mathrm{\&sigma;}}^2\left(m\right)}\right)---\left(13\right)$

The structure of gain function

The basic thought of Wiener filtering noise reduction model is by designing linear filter H (w) so that can reach mean square error expectation by the output after this linear filter minimum, namelyOptimal estimation for clean speech S (w).

$\hat{S}\left(w\right)=H\left(w\right)X\left(w\right)---\left(14\right)$

Wherein,P_sW () is clean speech signal power spectral density, P_dW () is noise power spectral density. Noise Estimation is generally adopted the noise spectrum estimation algorithm based on prior weight and posteriori SNR, and derives filter transfer function H (w) by introducing smoothing parameter.

Wiener Filter Method is by noisy speech being multiplied by a gain function or by a wave filter, its purpose is to the sound pressure level of noise attentuation to patient comfort. Therefore embodiment does not carry out noise abatement to realize the signals and associated noises to quiet signal or high signal noise ratio or only carries out slight noise abatement, and the noise cancellation signal that has of low signal noise ratio is improved reduction, is defined as by gain function

g_dB(i, m)=λ_im(SNR)f(x_dB(n))(15)

$g(i,m)={10}^{2\&CenterDot;g_{dB}(i,m)}---\left(16\right)$

Wherein g_db(i, m) represents the gain (dB territory) of the i-th frame voice signal in m subband, and (i is m) by g to g_db(i m) transforms to amplitude domain. λ_im(SNR) represent in m subband that the i-th frame voice signal is with the SNR pad value changed, by λ_im(SNR) it is limited in [-1,0]. Work as λ_im(SNR), when=0, being transformed into amount of gain after amplitude domain is 1, represents undamped; λ_im(SNR), time closer to-1, the pad value in respective amplitude territory is more big.

Although existing Noise reduction algorithm utilizes pad value function lambda_im(SNR) carry out noise reduction with the relation of subband SNR, but under low signal-to-noise ratio, background noise still remains too much, can cause user's auditory fatigue. For this, embodiment proposes a kind of improvement strategy, in low signal-to-noise ratio section modified gain proportion function, as it is shown on figure 3, low signal-to-noise ratio section [0, B0] attenuation curve is precipitous, it is therefore an objective to significantly weaken background noise during low signal-to-noise ratio. But this strategy is disadvantageous for faint quiet signal, and therefore gain function should also depend on environmental noise level. When being in quiet indoor such as patient, if relying solely on signal noise ratio calculating gain can cause that faint quiet signal is significantly attenuated, introduce maximum noise attenuation function f (N for this_dB(n)) in order to control signal attenuation amplitude. So under quiet environment, N_dBN () is very low, arrange maximum attenuation amplitude f (N_dB(n))=N_dBAfter (n), even if very low also the need hardly of signal noise ratio is decayed.

Performance test analysis

Compare the algorithm of present invention proposition and the algorithm performance difference of the patent US2004/6757395B1 modulation depth method proposed by experiment, and compare the performance of processing delay with basic spectrum-subtraction, Wiener Filter Method. Test all clean speech and take from TIMIT speech database, sample rate 16kHz. Pure noise intercepts from Noise92 noise storehouse, and noise type respectively White, Pink, Tank and SpeechBabble, input signal-to-noise ratio takes 0dB, 5dB, 10dB.

Speech perceptual quality is evaluated

Embodiment adopts logarithmic spectrum to estimate LSD (Log-spectraldistortion) and speech perceptual quality evaluation PESQ (Perceptualevaluationofspeechquality) is objective evaluation index.

LSD reflection is voice distortion situation, and PESQ will reflect voice quality on the whole. Both indexs and subjective assessment have higher degree of association. General LSD drop-out value (Δ LSD) is more big, it was shown that the logarithmic spectrum distortion factor is more little, and algorithm is more little to the degree of injury of voice, and PESQ raising amount (Δ PESQ) is more big, it was shown that the voice quality after algorithm process is more good. The computational methods of LSD and PESQ are as follows

$LSD=\frac{1}{J}\underset{l=0}{\overset{J-1}{\&Sigma;}}{\left\{\frac{1}{N/2+1}\underset{k=0}{\overset{N/2}{\&Sigma;}}{\&lsqb;l0{\log }_{10}X(k,l)-10{\log }_{10}\hat{X}(k,l)\&rsqb;}^2\right\}}^{\frac{1}{2}}---\left(17\right)$

PESQ=4.5-0.1 d_SYM-0.0309·d_ASYM(18)

In formula (17), X (k, l) andThe respectively Short Time Fourier Transform of clean speech and enhanced voice, N is frame length, and J is frame number. D in formula (18)_SYMAnd d_ASYMFor the correlometer formula in cognitive model. Result of calculation is as shown in table 3.

The speech perceptual quality evaluation of table 3 modulation depth method and the inventive method contrasts

In table 3, relatively input signal-to-noise ratio respectively 0dB, 5dB, during 10dB, the raising amount of the modulation depth method method proposed from embodiment LSD slippage under different noise types and PESQ, it appeared that the method that embodiment proposes shows excellence under White and Pink noise circumstance, the improvement degree of LSD and PESQ is substantially better than modulation depth method, particularly under White noise circumstance, the inventive method relatively modulation depth method on average exceeds 36.7% in logarithmic spectrum distance improved values, and PESQ improved values on average exceeds 19%. But under Tank and SpeechBabble noise circumstance, the present invention and modulation depth method performance are all performed poor, the average improvement amount of PESQ is only 0.1, so the linear scale attenuation model used by the present invention is further improved, be not suitable for Tank and SpeechBabble noise circumstance, therefore the performance of subjective feeling is by the impact of noise scenarios.

Algorithm complex and time delay analysis

The noisy speech duration 2.26s wherein processed, experiment frame length takes 64,128 and 256 respectively, and frame moves and is 50%.Although being both needed to, by analysis filter, signal is divided into some subbands for multiband algorithm, but the amount of calculation of filter equalizer being not belonging in the scope that the present invention considers. Therefore, the process time delay that embodiment is added up be from signal by after resolution filter to the process before synthesis filter. As can be known from Fig. 4, the time delay of basic spectrum-subtraction and basic Wiener Filter Method to be significantly greater than method proposed by the invention, this is because this two class model is both needed to by fast Fourier transform (FastFourierTransform, FFT) frequency-domain analysis is carried out, and it is final by inverse fast Fourier transform (InverseFastFourierTransform, IFFT) reverting to time-domain signal, therefore amount of calculation can dramatically increase. And the present invention by means of FFT and carries out frequency-domain analysis when the estimation of subband signal to noise ratio, the calculating of FFT is only for estimator (13), and the cross-correlation function being directed to and the amount of calculation of mean-square value are all extremely low. Additionally the present invention need to not revert to the process of time-domain signal via IFFT. Therefore the computation complexity invented is lower than basic spectrum-subtraction and basic Wiener Filter Method. Engineering generally takes frame length 2⁸=256 points, now the more basic spectrum-subtraction of processing delay performance of the present invention improves 60.6%, and relatively Wiener Filter Method improves 40.7%.

The above is only the preferred embodiment of the present invention; it should be pointed out that, for those skilled in the art, under the premise without departing from the technology of the present invention principle; can also making some improvement and deformation, these improve and deformation also should be regarded as protection scope of the present invention.

Claims (7)

1. the digital deaf-aid noise-reduction method based on improvement subband signal-to-noise ratio (SNR) estimation, it is characterised in that including:

Adopt analysis filter bank that original signal resolves into several subbands, then calculate the cross-correlation function of adjacent two frame signals in each subband and respective mean-square value, estimate the signal to noise ratio of this subband;

Secondly, according to each subband gain of the signal-to-noise ratio computation estimated, and the subband signal obtaining revising that is multiplied with subband signal;

Finally, revised each subband signal is comprehensively obtained the voice after noise reduction process.

2. as claimed in claim 1 based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, it is characterised in that to specifically include following steps:

S1, microphone input signal x (n) to digital deaf-aid are AD converted, and by the digital signal framing x after conversion_i(n) (i=0,1,2 ... N-1), wherein N is frame length;

S2, pass through resolution filter: by every frame signal by multichannel 6 rank IIR analysis filter bank h_i(n) (i=0,1 ... M-1), M is port number, with c_iCarry out down-sampled for down-sampled coefficient, obtain M subband signal: x (i, m) (i=0,1,2 ... N-1) (m=0,1,2 ... M-1);

S3, subband signal-to-noise ratio (SNR) estimation: the signal-to-noise ratio (SNR) estimation value of m-th subband the i-th frame is

$SNR(i,m)=10log\left(\frac{\mathrm{\&alpha;}P(i,m)+(1-\mathrm{\&alpha;})P(i-1,m)}{{\mathrm{\&sigma;}}^2\left(m\right)}\right)---\left(12\right)$

In formula, (i, m), (i-1, m) for the pure voice signal power of the in m-th subband i-th and i-th-1 frame, σ for P for P²M () is the noise variance in m-th subband, α is smoothing factor, 0 < α < 1;

S4, yield value calculate: by the signal to noise ratio snr that estimates, (i m) brings gain function g into_dB(i, in m), and by g_db(i, m) transform to amplitude domain g (i, m).

S5, subband noise suppress: primary speech signal is carried out noise attentuation by the yield value of each subband by obtaining, and obtains enhanced voice signal: y_im(n)=g (i, m) x (i, m);

S6, subband signal comprehensively export: by the output signal y of M passage_im(n) (i=0,1 ... M-1) first with c_iCarry out a liter sampling for liter downsampling factor, be finally synthesizing voice signal y (n) obtaining output.

3. as claimed in claim 2 based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, it is characterised in that in step S3, subband signal-to-noise ratio (SNR) estimation particularly as follows:

The i-th and i-th-1 frame signal on S31, definition kth Frequency point is adjacent signals vector X_adjoin(i, k), calculate adjacent signals vector autocorrelation matrix R (i, k);

S32, utilize the L in each subband_mIndividual frequency spectrum estimates correlation matrix R (i, the E [X in m)²(i,m)]、E[X²(i-1, m)] and E [X (i-1, m) X (i, m)], E [X²(i,m)]、E[X²(i-1, m)] mean-square value of respectively the i-th frame in m-th subband and the i-th-1 frame, E [X (i, m) X (i-1, m)] represent the i-th frame and the cross-correlation function of the i-th-1 frame;

S33, solve the i-th and i-th-1 frame in adjacent two frame voice signals and m-th subband pure voice signal power P (i, m), P (i-1, m), and the noise variance σ in m-th subband²(m), for estimate m-th subband the i-th frame signal-to-noise ratio (SNR) estimation value SNR (i, m).

4. as claimed in claim 3 based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, it is characterised in that in step S31, the autocorrelation matrix calculating adjacent signals vector is as follows:

$R(i,k)=E\&lsqb;X_{adjoin}(i,k)X_{adjoin}{(i,k)}^H\&rsqb;=\left(\begin{array}{cc}E\left(X^2\right(i,k\left)\right)& E\left(X\right(i,k\left)X\right(i-1,k\left)\right)\\ E\left(X\right(i,k\left)X\right(i-1,k\left)\right)& E\left(X^2\right(i-1,k\left)\right)\end{array}\right)---\left(1\right);$

Wherein, adjacent signals vector X_adjoin(i, k)=X (i, k), X (i-1, k) }^T, E (X²(i, k)), E (X²The i-th frame in (i-1, k)) respectively kth Frequency point kth subband and the mean-square value of the i-th-1 frame, ((i, k) x (i-1, k)) represents the i-th frame and the cross-correlation function of the i-th-1 frame to X to E.

5. as claimed in claim 3 based on the digital deaf-aid noise-reduction method improving subband signal-to-noise ratio (SNR) estimation, it is characterised in that in step S32, after signal wave filter by analysis obtains M subband, to comprise L according to each subband_mIndividual frequency spectrum estimate m-th subband consecutive frame correlation matrix R (i, m):

$R(i,m)=\left(\begin{array}{cc}E\&lsqb;X^2(i,m)\&rsqb;& E\&lsqb;X(i,m)X(i-1,m)\&rsqb;\\ E\&lsqb;X(i,m)X(i-1,m)\&rsqb;& E\&lsqb;X^2(i-1,m)\&rsqb;\end{array}\right)---\left(2\right)$

In formula (2), E [X²(i,m)]、E[X²(i-1, m)] mean-square value of respectively the i-th frame in m-th subband and the i-th-1 frame, E [X (i, m) X (i-1, m)] represent the i-th frame and the cross-correlation function of the i-th-1 frame.

6. the digital deaf-aid noise-reduction method based on improvement subband signal-to-noise ratio (SNR) estimation as described in any one of claim 1-5, it is characterised in that in step S4, gain function is

g_dB(i, m)=λ_im(SNR)f(x_dB(n))(13)

Wherein, g_db(i m) represents the gain of the i-th frame voice signal in m subband.

7. the digital deaf-aid noise-reduction method based on improvement subband signal-to-noise ratio (SNR) estimation as described in any one of claim 1-5 1, it is characterised in that in step S4, by the gain g of the i-th frame voice signal in m subband_db(i, m) transform to amplitude domain g (i, m):

$g(i,m)={10}^{2\&CenterDot;g_{dB}(i,m)}---\left(14\right).$