CN1967659A - Speech enhancement method applied to deaf-aid - Google Patents

Abstract

The invention provides a speech enhancement method that can restrain noise component in noisy voice and improve voice quality and intelligibility in noise environment. According to the method of the present invention, it specifically includes the following steps: 1) first, estimating the noise spectrum of the speech using endpoint detection technology; 2) removing noise using two-step Wiener filter technology; 3) meanwhile, adjusting the filter parameters according to auditory masking curve. The de-noising method in the invention has undertaken a thorough study and consideration on the signal level and perceptual level. For the estimation of noise, it uses the endpoint detection method, which has good robustness, to accurate estimate the noise spectrum as far as possible. Furthermore, it uses a two-step Wiener filter to solve this problem. Meanwhile, taking into account people's ears, by using auditory masking curve, it decreases the signal distortion. Thus, the eventually enhanced signal has a big increase on the voice quality and intelligibility.

Description

The sound enhancement method that is used for osophone

Technical field

The invention belongs to field of voice signal, relate to the sound enhancement method of osophone, be specifically related to a kind of method that in osophone, improves voice quality and the intelligibility of speech under the noise circumstance.

Background technology

Verbal communication is the basic communication mode of human society, also is one of basic viability of individual.Yet for those deafness patients, the verbal communication obstacle that causes owing to auditory dysesthesia has had a strong impact on their quality of life.This has brought huge misery not only for he or she and family, has also increased burden to entire society simultaneously.A statistics of announcing on February 7th, 2002 according to China Disabled Federation shows that there is the disabled person 2,057 ten thousand of hearing disfluency in China, accounts for 34.3% of the whole nation 6,000 ten thousand disabled person's sums.In addition, owing to reasons such as heredity, medicine, infection, noise, mishaies, also can increase 30,000 of deaf youngsters every year newly.So numerous barrier crowd and their life miseries of listening are being impelled the worker of association area to remove to use up portion and are being made great efforts to help these deaf persons and come back to the sound world, live as the normal person, embody humanity love of our harmonious society with this.

At present, the phonosensitive nerve hearing loss does not still have desirable conservative healing means, and main therapeutic intervention method is to wear osophone and implantable artificial cochlea, is suitable for Most patients and wear osophone.Make a definite diagnosis hearing rehabilitation from hearing loss, osophone is a most important and indispensable link in the whole chain.

Early stage osophone is analog machine, up to nineteen ninety-five, the first item digital hearing aid just occurred.In recent years, growing along with voice signal digital processing method and integrated circuit technique, osophone just progressively enters digital Age by the simulation epoch.Digital deaf-aid has overcome simple, the single defective of simulation function of hearing aid, can comparatively effectively distinguish target voice and interference noise, and might take the corresponding signal process strategy, thereby reached preliminary intellectuality by analyzing different application scenarioss.Digital deaf-aid is more and more accepted by the patient with its powerful signal handling capacity.

Yet under noise circumstance, the osophone wearer has serious decline to the intelligibility relative hearing normal person of voice, and after the amplification of noise process method, makes the wearer feel quite tired and bored.Current osophone method still can not solve this noise problem preferably, therefore, provides a kind of sound enhancement method that is used for osophone, has important Practical significance.

Summary of the invention

The purpose of this invention is to provide a kind of noise contribution of in noise circumstance, restraining in the noisy speech, improve the method for voice quality and intelligibility.

The concrete three aspect problems that solve of the present invention: the one, how exactly estimating noise is composed; The 2nd, how to adjust the noisy speech spectrum and suppress noise contribution; The 3rd, how to utilize sense of hearing mechanism of perception, make sound more natural, obtain better intelligibility.

The sound enhancement method according to the present invention specifically may further comprise the steps:

1) at first utilize the end-point detection technology that the noise spectrum in the voice is estimated;

2) utilize two step Wiener filtering technology to remove noise;

3) according to the auditory masking curve filtering parameter is adjusted simultaneously.

Describe the particular content of each technical step below in detail:

1. estimate based on the noise spectrum of end-point detection technology.

Adopt a kind of reliable among the present invention, to the reasonable end-point detection technology of noise robustness (VoiceActivity Dection-VAD), noise spectrum is upgraded being judged as in the time period of noise spectrum, and detect to contain in the time period of voice, its noise spectrum adopts the estimated value of front.For the spectrum of estimating noise as far as possible exactly, require in less voice being judged as under the situation of noise as much as possible noise segment that detect more.

The differentiation parameter that general end-point detecting method adopts is the energy of signal, zero-crossing rate, signal period property and linear forecast coding coefficient.They are for noisy situation poor-performing, and cutting is meticulous.

In order to realize the desired silence detection effect of native system, the end-point detecting method of voice messaging when we adopt based on length.Since this method information during based on voice signal long, big pause during it can estimate more exactly and speak, and mistakenly that energy is the less voiceless sound of minimizing is judged as the situation of noise segment.

2. two go on foot Wiener filterings.

This technology is used to adjust the noisy speech spectrum.The first step in the two step filtering is used for removing roughly noise, the albefaction residual noise; Second step was used to remove remaining white noise, obtained clean speech.

Input signal is composed estimation (Spectrum Estimation) behind minute frame, be averaged with the former frame result that to carry out time domain level and smooth again, obtains power spectrum density average (PSD Mean).The output result of these two modules is input in the design module (Weiner Filter Design) of S filter.In addition, import simultaneously this module to also have current frame signal be voice or quiet judged result.If quiet, this frame just can be used for estimating noise intensity.According to these input informations, wiener filter design is created on the coefficient on the linear frequency domain.Further, can obtain filter coefficient on the U.S. scale frequency axle by the conversion that beautifies the band module.Again by the time domain shock response of U.S. scale inverse discrete cosine transform (IDCT) module output filter.With the noisy speech signal of this response convolution input, can obtain the time domain speech waveform after the preliminary denoising.

The input of secondary filtering is exactly the output of a filtering, and processing procedure is similar to a filtering.Difference is that Noise Estimation partly improves.In addition, the U.S. scale filter coefficient that design is generated will be multiplied by a gain factor based on signal to noise ratio (S/N ratio) (Gain Factorization).

Final system is removed the direct current biasing among the secondary filtering result, and outputs to next unit.

3. based on the noise removing of auditory masking.

This technology is considered sense of hearing mechanism of perception on frequency spectrum processing when removing noise, make filter effect better.Concrete, in two Wiener filtering stages in step, all consider the influence of auditory masking curve, when the noise of estimating is being sheltered below the curve, no longer this part noise is handled, with the voice signal distortion that reduces to cause owing to filtering.

Advantage of the present invention and good effect:

Denoising method among the present invention has all been carried out sufficient research and consideration at signal level and sense of hearing perception level.For estimating noise, adopt the excellent in robustness end-point detecting method, that uses as much as possible accurately estimates noise spectrum.Consider simultaneously so still can not entirely accurate the value that obtains noise spectrum, and adopted two step Wiener filterings further to address this problem.Simultaneously, consider the characteristics of people's ear, utilization auditory masking curve, the distorted signals that obtains reduces.Like this, the signal behind final the enhancing all has bigger raising on voice quality and intelligibility.Use the osophone of this method, the usability under noise circumstance is significantly promoted.

Description of drawings

Below in conjunction with accompanying drawing the present invention is illustrated in further detail:

Fig. 1 is the overall system block diagram;

Fig. 2 is the end-point detecting method process flow diagram;

Fig. 3 is the FB(flow block) of two step Wiener filtering methods;

Fig. 4 is that signal compares with the result that PESQ carries out MOS marking after former signals and associated noises and the enhancing.

Embodiment

Below with reference to accompanying drawing of the present invention, best enforcement of the present invention described in more detail.

According to the sound enhancement method based on sense of hearing perception of the present invention, the general structure block diagram of system as shown in Figure 1, below with reference to the detailed introducing system structure of accompanying drawing.

1. the realization of end-point detection (VAD):

Characteristic parameter used in the present embodiment is defined as follows:

Suppose that x (n) is a noisy speech signal behind the branch frame, X (k is the amplitude spectrum of its l frame, k band l), and spectrum envelope (Long-Term Sprctral Envelop-LTSE) was defined as when the N rank were long so:

$\mathrm{LTSE}(k,l)=\max {\left\{X\right(k,l+j\}}_{j=-N}^{j=+N}$

And the N rank between voice and the noise when long spectrum degree of separation (Long-Term Spectral Divergence-LTSD) be defined as:

$\mathrm{LTSD}\left(l\right)={10\log }_{10}\left(\frac{1}{\mathrm{NFFT}}\underset{k=0}{\overset{\mathrm{NFFT}-1}{\mathrm{\Σ}}}\frac{{\mathrm{LTSE}}^2(k,l)}{N^2\left(k\right)}\right)---\left(2\right)$

The spectrum degree of separation is the characteristic parameter of difference voice and the suitable robust of non-voice when long.

Be illustrated in figure 2 as the process flow diagram of end-point detecting method, we are as can be seen from figure, the calculating of spectrum degree of separation is that spectrum envelope obtains when long when long, and the judgement of end-point detection only need be when long the size of spectrum degree of separation and the comparison of thresholding, so the setting of thresholding is quite important.It is to carry out self-adaptation by estimated noise spectrum size of coming out to adjust.

2. the realization of two step Wiener filterings:

Fig. 3 shows in the sound enhancement method FB(flow block) of two step Wiener filtering methods, below with the realization of each module in the concrete introduction method:

1) estimation module

Spectrum is estimated overlapping minute frame of input signal, and every frame 25ms is long, and frame moves 10ms, and signal divides frame by Hanning window.By the fast Fourier transform (FFT) conversion, be transformed into frequency domain, and obtain power spectrum then, power spectrum is carried out smoothing processing.

2) power spectrum density average

Front and back two frame power spectrum densities are asked on average, made frequency spectrum more level and smooth in time.

3) wiener filter design

We consider to carry out two step Wiener filterings, in the Noise Estimation of first step filtering stage, are the results by end-point detection, upgrade in non-speech segment.In the Noise Estimation of the second step filtering stage, utilize the correlativity between voice and the noise to upgrade.

Priori SNR is:

$\mathrm{\η}(\mathrm{bin},t)=\frac{P_{\mathrm{den}}(\mathrm{bin},t)}{P_{\mathrm{notse}}(\mathrm{bin},t)}---\left(3\right)$

Filter transfer function H (bin t) is obtained by following equation:

$H(\mathrm{bin},t)=\frac{\sqrt{\mathrm{\η}(\mathrm{bin},t)}}{1+\sqrt{\mathrm{\η}(\mathrm{bin},t)}}---\left(4\right)$

4) U.S. scale bank of filters

Linear frequency is mapped to U.S. scale frequency, and for the 16k sampled signal, we are divided into 25 bands with whole frequency band, and linear frequency is transformed into U.S. scale frequency, calculate quarter window then, original Wiener filtering frequency response quarter window weighting on frequency.Obtain the Wiener filtering response under the U.S. scale like this.

5) gain factor

Gain factor is to be used in second Wiener filtering stage in the step adjustment squelch intensity.Gain factor is used on the U.S. scale S filter of being obtained:

H _{2_mel_GF}(k，t)＝(1-α _GF(t))+α _GF(t)×H _{2_mel}(k，t)，0≤k≤K _FB+1(5)

The variation range of gain factor from 0.1 to 0.8 means that to containing voice segments, filtering strength is lowered to 10%, and is set at 80% for noise segment in the second Wiener filtering stage in step.

6) U.S. scale IDCT

Calculate filter response from the frequency domain transform to the time domain with IDCT:

$h_{\mathrm{WF}}\left(n\right)=\underset{k=1}{\overset{K_{\mathrm{FB}}+1}{\mathrm{\&Sigma;}}}{\mathrm{<figure-callout\; id="2"\; label="H"\; filenames="A20051008687700102.png"\; state="\{\{state\}\}">H</figure-callout>}}_{2\_\mathrm{mel}}\left(k\right)\&times;{\mathrm{IDCT}}_{\mathrm{mel}}(k,n),0\&le;n\&le;K_{\mathrm{FB}}+1---\left(6\right)$

Wherein (k n) is U.S. scale IDCT to IDCTmel.Like this, we have just obtained time domain filtering.

7) filtering

From top time domain filtering, we calculate the cause and effect impact response filter, with input signal process wave filter, the signal after can being enhanced.

8) migration

In order to remove filtered DC component, the signal after the denoising is passed through following wave filter:

s _{nr_of}(n)＝s _nr(n)-s _nr(n-1)+(1-1/1024)×s _{nr_of}(n-1)，0≤n≤M-1(7)

3 filtering adjustment based on auditory masking

We are when carrying out two step Wiener filterings, and for the transition function of each frequency, we consider that also noise is whether below the auditory masking curve.If like this, the gain with this frequency is adjusted into 1 so, promptly it is not carried out denoising.Can reduce the waveform distortion that causes owing to denoising like this.

Signal compares with the result that PESQ carries out MOS marking after Figure 4 shows that former signals and associated noises and enhancing, signal after wherein curve 1 strengthens, curve 2 expression original signals more as can be seen, obviously have better effect through the signal after the enhancing according to the result of the marking of the MOS among the figure.

Although disclose specific embodiments of the invention and accompanying drawing for the purpose of illustration, its purpose is to help to understand content of the present invention and implement according to this, but it will be appreciated by those skilled in the art that: without departing from the spirit and scope of the invention and the appended claims, various replacements, variation and modification all are possible.Therefore, the present invention should not be limited to most preferred embodiment and the disclosed content of accompanying drawing.

Claims (7)

1. sound enhancement method that is used for osophone specifically may further comprise the steps:

1) utilize the end-point detection technology that the noise spectrum in the voice is estimated;

2) utilize two step Wiener filterings to remove noise;

3) according to the auditory masking curve filtering parameter is adjusted simultaneously.

2. the sound enhancement method that is used for osophone as claimed in claim 1 is characterized in that: the end-point detection technology of voice messaging is estimated the noise spectrum in the voice when adopting based on length.

3. the sound enhancement method that is used for osophone as claimed in claim 1 or 2 is characterized in that: utilize the end-point detection technology that the noise spectrum in the voice is estimated, in the time period that is judged as noise spectrum noise spectrum is upgraded; When detecting to contain in the time period of voice, its noise spectrum adopts existing estimated value.

4. the sound enhancement method that is used for osophone as claimed in claim 1 is characterized in that:

First step Wiener filtering is removed noise, and at first input signal is composed estimation behind minute frame, is averaged with the former frame result that to carry out time domain level and smooth again, obtains the power spectrum density average; Then these two output results are input in the S filter module;

To be that voice still are quiet judged result also simultaneously, be input to the S filter module current frame signal;

If quiet, this frame just can be used for estimating noise intensity;

According to the input information of S filter module, wiener filter design is created on the coefficient on the linear frequency domain;

Further, the conversion by U.S. wave filter band module can obtain the filter coefficient on the U.S. scale frequency axle; Again by the time domain shock response of U.S. scale inverse discrete cosine transform module output filter;

With the noisy speech signal of this response convolution input, obtain the time domain speech waveform after the preliminary denoising.

5. the sound enhancement method that is used for osophone as claimed in claim 4, it is characterized in that: the input of the second step Wiener filtering is the output of first step Wiener filtering, and the processing procedure difference is that the correlativity between Noise Estimation part voice and the noise upgrades; And the U.S. scale filter coefficient that design is generated multiply by a gain factor based on signal to noise ratio (S/N ratio).

6. the sound enhancement method that is used for osophone as claimed in claim 1 is characterized in that: in two Wiener filtering stages in step, all consider the influence of auditory masking curve, when the noise of estimating is being sheltered below the curve, no longer this part noise is handled.

7. the sound enhancement method that is used for osophone as claimed in claim 6 is characterized in that: when the noise of estimating is being sheltered below the curve, the gain of this frequency is adjusted into 1.

Images