CN100524465C - A method and device for noise elimination - Google Patents

Abstract

The invention discloses a noise elimination device and method. Two microphones are used to construct a heart-shaped directional microphone with two zero points respectively pointing to the front and rear directions, so that one signal of the main microphone after processing mainly contains the target voice, and the other signal mainly contains noise. ; Then use frequency-domain adaptive filtering to control the beam pointing range of the entire device by adjusting the gain of the adaptive filter; finally, use a single-channel denoising algorithm to further eliminate the noise components in the main microphone signal. This allows the device to achieve optimum performance in terms of noise cancellation and maintaining target voice quality. The invention can be widely used in the field of microphone noise elimination.

Description

Noise canceling device and method

technical field

The invention relates to the field of audio processing, in particular to a noise elimination device and method.

technical background

In addition to the target voice signal, the sound received by the microphone often also includes noises such as background vocals and background music. The methods of noise elimination in the prior art are mainly divided into two methods: single-microphone de-noising and multi-microphone, that is, microphone array de-noising.

The single-microphone denoising method can achieve better results in specific application scenarios. The denoising principle generally uses the difference between the target speech and noise components in the time-frequency domain to eliminate noise. For example, it is generally believed that the noise signal The characteristics change relatively slowly for speech signals. Therefore, for the general case, this algorithm has the following disadvantages:

1) When the noise is non-stationary type, such as background vocals and background music, the denoising effect is relatively lower than that of the stationary type noise;

2) When the signal-to-noise ratio is low (such as lower than 0), the denoising effect is not obvious;

3) Music noise may be introduced, and the music noise is in the noise elimination process, so a compromise needs to be made between maintaining voice quality and suppressing noise.

The multi-microphone denoising method can overcome the above problems to adapt to the harsh acoustic environment, so it has been widely used in digital hearing aids, vehicle voice equipment or recording equipment, communication devices for pilots and soldiers, conference microphones, and voice recognition front-end devices. Applications. Compared with the single-microphone denoising method, the multi-microphone denoising method mainly utilizes the difference in the spatial domain between the target sound source and the noise source, that is, the distance and direction of each sound source to the microphone are different to separate the signals, thereby achieving noise elimination.

The existing broadside type dual-microphone denoising device mainly has three methods: adopting two omnidirectional microphones; adopting two unidirectional microphones; adopting one omnidirectional and one unidirectional microphone. The so-called shoulder-to-shoulder mode refers to the time when the voice signal of the target directly in front reaches the two microphones at the same time, which is called broadside type.

The first method, that is, the method of using two omnidirectional microphones, often has a relatively large impact on the quality of the target voice, and basically cannot eliminate the noise from the front.

The second method is to use two forward-pointing unidirectional microphones, except that the characteristics of the unidirectional microphone itself can suppress the noise directly behind to a certain extent, and other performances are similar to the first method.

The third method is a combined method of a unidirectional microphone and an omnidirectional microphone. Existing solutions often directly use these two signals for adaptive filtering. Since the two signals contain strong voice components, the voice quality will deteriorate. has declined.

To sum up, the existing dual-microphone denoising method has an unsatisfactory effect of noise elimination, and while eliminating noise, it will cause a decrease in the quality of the target voice.

Contents of the invention

In view of the shortcomings of the side-by-side dual-microphone denoising method in the prior art, the purpose of the present invention is to provide a noise canceling device and method, which can effectively eliminate noise without causing degradation of the target voice quality.

In order to achieve the above object, the present invention provides a noise canceling device, comprising:

The target speech signal acquisition module is used to obtain the sound signal whose main component is the speech signal;

The noise signal acquisition module is used to acquire the sound signal whose main component is the noise signal;

Adaptive filtering module, for using the output signal simulation of the noise signal acquisition module to obtain the noise signal component in the target speech signal acquisition module, and then subtract the simulated signal from the output of the target speech signal acquisition module, To obtain a speech signal with noise removed;

The single-channel denoising module is configured to perform single-channel denoising processing on the output result of the adaptive filtering module to obtain a signal after further noise removal.

Preferably, the target voice signal acquisition module is a cardioid unidirectional microphone pointing in the direction of the target voice signal, and is used to absorb the sound signal whose main component is the target voice signal;

The noise signal acquisition module includes an omnidirectional microphone, a gain adjustment unit, and a subtractor, the omnidirectional microphone is used to absorb sound signals in all directions, and the gain adjustment unit is used to adjust the omnidirectional The gain of the directional microphone output signal makes the gain of the omnidirectional microphone and the cardioid unidirectional microphone absorbing the sound of the target voice direction the same, and the subtractor is used to combine the signal adjusted by the omnidirectional microphone with the gain The outputs of the cardioid-shaped unidirectional microphones are subtracted to obtain a sound signal whose main component is a noise signal.

Preferably, the target voice signal acquisition module is a cardioid unidirectional microphone pointing in the direction of the target voice signal, and is used to absorb the sound signal whose main component is the target voice signal;

The noise signal acquisition module is a unidirectional microphone whose direction is opposite to that of the target voice signal, and is used to absorb the voice signal whose main component is the noise signal.

Preferably, the adaptive filtering module is a frequency domain adaptive filter.

Preferably, the frequency-domain adaptive filter includes a coefficient adjustment unit for detecting the size of the filter coefficient, and reducing the value of the filter coefficient when the filter coefficient is too large.

Preferably, the value of the reduced filter coefficient is specifically: $W(k+1)\′=W(k+1)*\frac{\mathrm{Threshold}}{\left|\right|W_{\max }(k+1)\left|\right|},$

W(k+1)=[W ₀ (k+1), W ₁ (k+1)...W _N-1 (k+1)] is the coefficient of the frequency domain adaptive filter, which is an N Dimensional complex vector, N is the coefficient length;

‖W _max (k+1)‖ is the maximum value of the modulus of each complex number of W ₀ (k+1), W ₁ (k+1)...W _N-1 (k+1);

$\mathrm{Threshold}=\sqrt{\frac{1+\cos \left(\mathrm{\θ}\right)}{1-\cos \left(\mathrm{\θ}\right)}}$ Among them θ is the input angle value of the speech signal that wishes to protect.

The present invention also provides a noise elimination method, comprising the following steps:

Obtain the target speech signal, that is, the sound signal s(k) whose main component is a speech signal;

Obtain a noise signal, that is, a sound signal n(k) whose main component is noise;

The adaptive filtering step is to filter the n(k) to simulate the noise signal component in the s(k), and then subtract the simulated noise from s(k);

A single-channel denoising step, performing single-channel denoising processing on the result of the adaptive filtering step to obtain a signal with noise removed.

Preferably, the acquisition of the target voice signal is to use a unidirectional microphone to absorb the sound signal f(k) of the target voice direction, that is, s(k)=f(k);

The acquisition of noise signals includes:

Use an omnidirectional microphone to absorb sound signals b(k) from all directions;

Adjusting the gain α of the b(k) to make it the same as the gain of the f(k) in the target voice direction;

The signal α*b(k) after gain adjustment is subtracted from the f(k) to obtain the sound signal n(k) whose main component is a noise signal, i.e. n(k)=α*b( k)-f(k).

Preferably, the acquisition of the target voice signal is to use a unidirectional microphone to absorb the sound signal f'(k) of the target voice direction, that is, s(k)=f'(k);

The acquisition of the noise signal is to use a cardioid unidirectional microphone to absorb the sound signal b'(k) opposite to the direction of the target voice, that is, n(k)=b'(k).

Preferably, the adaptive filtering step adopts a frequency domain adaptive filtering method.

Preferably, the size of the filter coefficient is detected after each update of the filter coefficient, and when the filter coefficient is too large, its value is reduced.

Preferably, the value of the reduced filter coefficient is specifically: $W(k+1)\′=W(k+1)*\frac{\mathrm{Threshold}}{\left|\right|W_{\max }(k+1)\left|\right|},$

The W(k+1)=[W ₀ (k+1), W ₁ (k+1)...W _N-1 (k+1)] is the coefficient of the frequency domain adaptive filter, which is An N-dimensional complex vector, where N is the coefficient length;

The ∥W _max (k+1)∥ is the maximum value of the modulus of each complex number of the W ₀ (k+1), W ₁ (k+1)...W _N-1 (k+1);

said $\mathrm{Threshold}=\sqrt{\frac{1+\cos \left(\mathrm{\θ}\right)}{1-\cos \left(\mathrm{\θ}\right)}}$ Among them, θ is the maximum value of the input angle of the speech signal to be protected.

The difference between the present invention and the prior art is that two microphones are used to make the main microphone one-way signal after processing mainly contain the target voice, while the auxiliary microphone one-way signal mainly contains noise; then frequency domain adaptive filtering is adopted, and by adjusting The gain of the adaptive filter is used to control the beam pointing range of the entire device, so that the performance of the device for eliminating noise and maintaining the target voice quality is optimized. The invention can effectively eliminate the noise without reducing the quality of the target voice.

Description of drawings

Fig. 1: the basic principle block diagram of device of the present invention;

Fig. 2: the circuit block diagram of embodiment one of the present invention;

Fig. 3: the circuit block diagram of embodiment two of the present invention;

Figure 4: Schematic diagram of beamforming;

Figure 5: Schematic diagram of frequency domain LMS algorithm;

Figure 6: Omnidirectional microphone and cardioid unidirectional microphone polar diagram;

Figure 7: Schematic diagram of the method of the present invention.

Detailed ways

On the basis of the prior art, the present invention proposes a new side-by-side dual-microphone denoising device and method through sufficient theoretical research and a large number of experiments. The present invention uses two microphones to construct two cardioid-shaped microphones with zero points pointing to the front and rear directions respectively, so that one signal of the main microphone after processing mainly contains the target voice, and the other signal mainly contains noise. As mentioned in the article, pointing to the front means pointing to the direction of the target voice, and pointing to the back means pointing to the opposite direction of the target voice; one line of the main microphone refers to the one line of microphone that mainly collects the target voice signal, and one line of the auxiliary microphone refers to the signal obtained after processing. A microphone for the noise signal.

The device and method of the present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Fig. 1, it is a basic principle block diagram of the device of the present invention. It includes: a target speech signal acquisition module, which is used to obtain a sound signal whose main component is a speech signal;

The noise signal acquisition module is used to acquire the sound signal whose main component is the noise signal;

Adaptive filtering module, for using the output signal simulation of the noise signal acquisition module to obtain the noise signal component in the target speech signal acquisition module;

The single-channel denoising module is used to subtract the output of the target speech signal acquisition module from the output of the adaptive filtering module to obtain a speech signal from which noise has been removed.

Two exemplary embodiments of the device according to the invention are given here. The first embodiment, as shown in Figure 2, uses a cardioid unidirectional microphone to collect f(x) whose main component is the target speech signal, and simultaneously uses an omnidirectional microphone to collect the sound signal b(x), and then uses beamforming The module processes the f(x) and b(x) to construct a cardioid-shaped microphone whose zero points of signal gain point to the front and rear directions respectively. The second embodiment, as shown in FIG. 3 , directly uses two cardioid unidirectional microphones pointing to the front and rear directions respectively. Both of the above two methods can make the signal of one channel of the main microphone after processing mainly contain the target voice, and the signal of the other channel mainly contain noise. Then, a frequency-domain adaptive filter is used to control the beam pointing range of the entire device by adjusting its gain. Finally, the single-channel denoising module is used to further eliminate the noise components in the main microphone signal. This allows the device to achieve optimum performance in terms of noise cancellation and maintaining target voice quality.

In the two embodiments, there are different ways of constructing two cardioid microphones with the zero points of signal gain pointing to the front and rear directions respectively, and the construction principles are described in detail below.

As shown in Figure 4: the formation of the beam in the first embodiment needs to include a cardioid unidirectional microphone, and the characteristic curve of its gain is shown in the cardioid diagram on the right side of the unidirectional microphone in the figure, which points to the direction of the target voice, that is, the front The gain is the largest, and the gain for the rear sound signal is the smallest 0. An omnidirectional microphone is also included to collect sound signals from all directions with the same gain in all directions. The beamforming block shown in Figure 2 consists of a gain adjustment block and a subtractor in Figure 4. The gain adjustment module shown in the figure is used to adjust the gain of the output signal of the omnidirectional microphone, so that the gain of the omnidirectional microphone and the cardioid unidirectional microphone to absorb the sound of the target voice direction is the same, and the subtractor is used for The gain-adjusted signal of the omnidirectional microphone is subtracted from the output of the cardioid unidirectional microphone, and the difference obtained contains less target speech signals and mainly includes noise signals from other directions.

The mathematical expression of beamforming can be expressed as:

s(k)=f(k)

n(k)=α*b(k)-f(k) (1.2)

in:

f(k): the signal received by the cardioid unidirectional microphone pointing at the target voice (defined as 0 degree direction);

b(k): the signal received by the omnidirectional microphone;

s(k): The output of the constructed cardioid-shaped microphone whose zero point is 180 degrees. Theoretically, its main component is the target voice signal incident from the front.

n(k): The output of the constructed cardioid pointing microphone whose zero point is 0 degrees, theoretically its main component is the noise signal incident from the rear.

α: Gain factor, adjusted by the gain adjustment module, used to calibrate the microphones so that the two microphones have the same gain for absorbing the front signal.

After processing by this beamforming method, one signal of the main microphone mainly includes the target speech, while the signal of the other auxiliary microphone mainly includes noise. This provides ideal conditions for the next step of adaptive filtering.

In the second embodiment of the present invention, the signals collected by two cardioid-shaped unidirectional microphones pointing forward and backward respectively are equivalent to the two-way signals generated after beamforming processing in Embodiment 1, so beamforming is not required, and the Adaptive filtering. Specifically, use a heart-shaped unidirectional microphone to point to the front to absorb the sound signal f'(k) of the target voice direction, i.e. s(k)=f'(k); in addition, use a heart-shaped unidirectional microphone to absorb and target voice The sound signal b'(k) in the opposite direction, ie n(k)=b'(k).

The frequency domain adaptive filtering module is introduced below. The reason why the present invention adopts frequency-domain adaptive filtering is mainly to consider the following four points:

1) Frequency-domain adaptive filtering has low computational complexity and higher efficiency;

2) Frequency-domain adaptive filtering has better robust performance;

3) Frequency-domain adaptive filtering, with good frequency selection characteristics, can simultaneously eliminate the noise generated by multiple interference noise sources with different frequency components;

4) It is convenient to apply the method proposed by the present invention, and control the beam pointing range of the whole device by adjusting the gain of the adaptive filter.

Here, the frequency domain adaptive filtering method used in the present invention is briefly introduced. The LMS algorithm in the frequency domain is used here, as shown in FIG. 5 , where the thin arrows represent the time domain signal processing, and the thick arrows represent the frequency domain signal processing. Adaptive filtering in the frequency domain is used, and the signal is processed in frames. We know that after the long sequence is truncated and processed in blocks and then merged, it is necessary to use the overlap-add method or the overlap-preserving method to avoid aliasing. This paper adopts the overlap-preserving method.

First, assume that the order of the adaptive filter we use is M, and its time-domain filter coefficient is recorded as w(k). Because the overlap-preserving method is used, in order to avoid aliasing, the M-order filter is extended with M 0s. A filter with N=2M coefficients is formed, and the frequency-domain coefficient vector of the filter obtained after FFT processing is:

$\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; FFT</span>}\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.1</span>}<span\; class="notranslate">\; )</span>$

It can be seen from the above formula that the length of the frequency domain adaptive filter coefficient vector is twice that of the time domain coefficient vector. For the frequency domain adaptive filtering algorithm, both adaptive filtering and filter coefficient update are completed in the frequency domain, So there will be no form of time-domain filter. It is worth noting that the FFT or inverse FFT we mentioned later are all N-point FFTs.

即为上文中所述的主要成分为噪声信号的n(k)经分帧后的数据，每帧数据长度为M，将上一帧

和当前帧

合并为一个大帧

如下所示：Next we consider the input signal, in the following description of the adaptive filtering method in the frequency domain, the

It is the framed data of n(k) whose main component is the noise signal mentioned above, the data length of each frame is M, and the previous frame

and the current frame

Merge into one large frame

As follows:

为合并后的大帧，长度为N＝2M。in

It is a merged large frame with a length of N=2M.

做FFT，转换到频域有：Will

Do FFT, convert to frequency domain:

$\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; FFT</span>}<span\; class="notranslate">\; [</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.3</span>}<span\; class="notranslate">\; )</span>$

Then we use the overlap preservation method to filter the input signal, that is, convolution in the time domain, or multiplication in the frequency domain, that is:

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; IFFT</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.4</span>}<span\; class="notranslate">\; )</span>$

Among them, the results of IFFT take the last M results,

表示上文中所述的主要成分为语音信号的s(k)：use here

Indicates that the main component described above is s(k) of the speech signal:

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.5</span>}<span\; class="notranslate">\; )</span>$

Then the filtered result signal is:

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; kM</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>$

(2.6)

$<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>$

After FFT, the error signal vector in the frequency domain is obtained as:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; FFT</span>}\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; 0</span>}\\ \stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.7</span>}<span\; class="notranslate">\; )</span>$

Similar to the time-domain LMS, now we calculate the update amount of the adaptive filter coefficient vector through the error signal vector E(k) and the input signal vector. In the frequency domain, the update amount of the adaptive filter coefficient vector is determined by calculating the correlation between the error signal and the input signal. Since the linear correlation is equivalent to an inverse linear convolution in form, with the help of the time domain The convolution of has a fast FFT algorithm in the frequency domain. According to the overlap-preserving method, there is

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; IFFT</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.8</span>}<span\; class="notranslate">\; )</span>$

The IFFT result takes the first M values.

来更新自适应滤波器系数，注意到频域的滤波器系数是将时域系数后面补零，然后经FFT处理生成的。所以相应的，我们就得到了滤波器系数更新的频域形式W(k+1)如下：Finally we use

To update the adaptive filter coefficients, note that the filter coefficients in the frequency domain are generated by padding the time domain coefficients with zeros and then undergoing FFT processing. So correspondingly, we get the frequency domain form W(k+1) of filter coefficient update as follows:

$\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&mu;FFT</span>}\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.9</span>}<span\; class="notranslate">\; )</span>$

Among them, μ is the step size of the filter.

It can be seen from the above description that the function of adaptive filtering is to make the signal n(k) of the auxiliary microphone after beamforming, through the filtering of the adaptive filter, simulate the noise signal in the main microphone. In this way, the noise signal in the main microphone is further eliminated.

However, since the characteristics of the microphone are not ideal, the signal of the auxiliary microphone contains a certain target voice signal. Therefore, if the adaptive filter is not controlled, a part of the voice signal may be eliminated, which will cause voice quality degradation. At present, many existing algorithms have this problem.

This paper proposes to control the beam pointing range of the entire device by adjusting the gain of the adaptive filter, so that the sound within the beam pointing range of the device will not be weakened by the adaptive filter, so that the target voice quality will not be degraded.

The reason why the gain control method can be adopted is that, through the two beam-formed signals, the target voice signal of one channel of the main microphone is much larger than the target voice signal of one channel of the auxiliary microphone. If the adaptive filter is trying to eliminate the target speech in the main microphone, the magnitude of the filter coefficients needs to be relatively large, that is to say, the adaptive filter needs a large gain. If we limit the gain of the adaptive filter within a threshold, then the adaptive filter cannot eliminate the target speech.

The method used is to check the size of the coefficient after each coefficient update. If it is greater than the set threshold, we think that the adaptive filter is trying to eliminate the target speech. Therefore, the gain of the filter is reduced to protect the target voice quality. Specifically, for the frequency-domain NLMS algorithm, as mentioned earlier, the coefficient update is as follows:

$\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&mu;FFT</span>}\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.9</span>}<span\; class="notranslate">\; )</span>$

In the formula, W(k+1) is the updated frequency-domain adaptive filter coefficient, which is an N-dimensional complex vector, and N is the number of FFT points, that is, W(k+1)=[W ₀ (k+1), W ₁ (k+1), ..., W _N-1 (k+1)] ^T (2.10)

The size of the coefficient is measured by the modulus of the complex number, namely:

[‖W ₀ (k+1)‖,‖W ₁ (k+1)‖,...,‖W _N-1 (k+1)‖] ^T (2.11)

For ‖W _i (k+1)‖, i=0, 1,..., N-1, search to find the modulus ‖W _max (k+1)‖ of the largest coefficient, if ‖W _max (k+1)‖ >Threshold, it is determined that the adaptive filter is trying to eliminate the target speech at this time, so the gain of the filter is reduced, that is: $W(k+1)\&prime;=W(k+1)*\frac{\mathrm{Threshold}}{\left|\right|W_{\max }(k+1)\left|\right|}.$

The selection of the Thrashold value is introduced below.

First refer to the cardioid unidirectional microphone and the omnidirectional microphone polar diagram used in Embodiment 1 of the present invention, as shown in Figure 6: wherein the polarity of the omnidirectional microphone does not change with the angle, while the cardioid unidirectional The microphone, which is the largest at an angle of 0 degrees and the smallest at an angle of 180 degrees, can be expressed in mathematical expressions as follows:

表示心形单指向性麦克风的极性。 $\left\{\begin{array}{c}{P}_{\mathrm{omni}}=1\\ P_{\mathrm{uni}}=0.5(1+\cos \mathrm{\&theta;})\end{array}\right.0\&le;\mathrm{\&theta;}<360,$ where _Pomm represents the polarity of the omnidirectional microphone, and

Indicates the polarity of a cardioid unidirectional microphone.

Then, the polarity ratio of the main microphone path and the auxiliary microphone path after beamforming is

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; uni</span>}}}{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; omni</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; uni</span>}}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&theta;</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&theta;</span>}}<span\; class="notranslate">\; ,</span>$

For the second embodiment, since two cardioid-shaped unidirectional microphones are used, the polarities are opposite, that is, the pointing angles differ by 180 degrees, which can be expressed as follows in a mathematical expression:

$\left\{\begin{array}{c}{P}_{\mathrm{uni}\_\mathrm{ref}}=0.5(1+\cos (\mathrm{\&theta;}+180))\\ P_{\mathrm{uni}}=0.5(1+\cos \mathrm{\&theta;})\end{array}\right.0\&le;\mathrm{\&theta;}<360,$ Where P _umi represents the polarity of the primary microphone, and P _{uni_ref} represents the polarity of the secondary microphone.

Then the polarity ratio of the main microphone path and the auxiliary microphone path is: $P\left(\mathrm{\&theta;}\right)=\frac{P_{\mathrm{uni}}}{P_{\mathrm{uni}}-P_{\mathrm{ref}}}=\frac{1+\cos \mathrm{\&theta;}}{1-\cos \mathrm{\&theta;}}.$

It can be seen that the following conclusions can be drawn from Examples 1 and 2, that is, P(θ) is maximum at 0 degrees, 0 at 180 degrees, and 1 at 90 degrees. That is, the energy contrast between the main microphone path and the auxiliary microphone path after beamforming, at 0 degrees, the former is much larger than the latter, at 180 degrees, the former is much smaller than the latter, and at 90 degrees, the two are almost the same. Therefore, we can determine Threshold according to P(θ), as shown in the following formula:

Threshold=P(θ)

For example, we want to protect the signal within the range of 0 degree angle as the center and 30 degree angle offset left and right. Then it can be found that: $\mathrm{Threshold}=P(\mathrm{\&pi;}/6)=\sqrt{\frac{1+\cos (\mathrm{\&pi;}/6)}{1-\cos (\mathrm{\&pi;}/6)}}=\&ap;3.73.$

There are mainly three methods for single-channel denoising: Wiener filtering, spectral subtraction and short-time spectrum adjustment. The single-channel denoising module of the present invention uses short-time spectrum adjustment to remove residual noise. There are many introductions in the literature, and their descriptions are omitted here.

The flow chart of the method of the present invention is shown in FIG. 7 , and its detailed content has been reflected in the aforementioned device introduction, so it will not be repeated here.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention within.

Claims (8)

1, a kind of noise elimination apparatus is characterized in that, comprising:

Target voice signal acquisition module is used to obtain the voice signal that the signal principal ingredient is a voice signal;

The noise signal acquisition module is used to obtain the voice signal that the signal principal ingredient is a noise signal;

The auto adapted filtering module, be used for the output signal of described noise signal acquisition module is simulated to obtain the noise signal composition of described target voice signal acquisition module, from the output of target voice signal acquisition module, deduct the described signal that obtains through simulation then, to obtain having removed the voice signal of noise;

Single channel denoising module is used for the output result of described auto adapted filtering module is carried out the single channel denoising, has further been removed the signal behind the noise,

Described target voice signal acquisition module is the heart-shaped unidirectivity microphone of a definite object voice signal direction, is used to absorb the voice signal that principal ingredient is the target voice signal;

Described noise signal acquisition module, comprise an omni-directional microphone and a gain adjusting unit, an and subtracter, described omni-directional microphone is used to absorb the voice signal of all directions, described gain adjusting unit is used to adjust the gain of described omni-directional microphone output signal, make that the gain of described omni-directional microphone and heart-shaped unidirectivity microphone absorption target voice direction sound is identical, described subtracter is used for described omni-directional microphone is subtracted each other by the signal of gain adjustment and the output of described heart-shaped unidirectivity microphone, obtains the voice signal that the signal principal ingredient is a noise signal; Perhaps, described noise signal acquisition module is a sensing and the reverse heart-shaped unidirectivity microphone of target voice signal direction, is used to absorb the voice signal that principal ingredient is a noise signal.

2, device according to claim 1 is characterized in that, described auto adapted filtering module is an adaptive frequency domain filter.

3, device according to claim 2 is characterized in that, comprises the coefficient adjustment unit in the described adaptive frequency domain filter, is used for the size of detection filter device coefficient, and when described filter coefficient is excessive, reduces its value.

4, device according to claim 3 is characterized in that, the value after described filter coefficient is lowered is: order $W(k+1)\&prime;=W(k+1)*\frac{\mathrm{Threshold}}{\left|\right|W_{\max }(k+1)\left|\right|},$

Described W (k+1)=[W ₀(k+1), W ₁(k+1) ... W _N-1(k+1)], be the coefficient of adaptive frequency domain filter, be the complex vector of a N dimension, N is a coefficient length;

Described W (k+1) ' is the value after described filter coefficient is lowered;

Described ‖ W _Max(k+1) ‖ is W ₀(k+1), W ₁(k+1) ... W _N-1The maximal value of the mould that (k+1) each is plural;

Described $\mathrm{Threshhold}=\sqrt{\frac{1+\cos \left(\mathrm{\&theta;}\right)}{1-\cos \left(\mathrm{\&theta;}\right)}},$ Wherein θ is the input angle value of the voice signal of hope protection.

5, a kind of dual microphone noise cancellation method is characterized in that, comprising:

Obtain the target voice signal, promptly principal ingredient is the voice signal s (k) of voice signal;

Obtain noise signal, promptly principal ingredient is the voice signal n (k) of noise;

The auto adapted filtering step is carried out filtering to described n (k), obtains noise signal composition among the described s (k) with simulation, makes s (k) deduct the noise that simulation obtains then;

Single channel denoising step is carried out the single channel denoising to the result of described auto adapted filtering step, has obtained removing the signal behind the noise,

The described target voice signal that obtains is for using the voice signal f (k) of unidirectivity microphone absorption target voice direction, i.e. s (k)=f (k);

The described noise signal of obtaining comprises: use an omni-directional microphone to absorb the voice signal b (k) of all directions; Adjust the gain alpha of described b (k), make it identical with the gain of described f (k) on the target voice direction; Subtract each other by adjusted signal alpha * b of gain (k) and described f (k) described, obtaining the signal principal ingredient is the voice signal n (k) of noise signal, i.e. n (k)=α * b (k)-f (k); Perhaps, the described noise signal of obtaining, for using a voice signal b ' that heart-shaped unidirectivity microphone absorbs and the target voice direction is reverse (k), promptly (k)=b ' is (k) for n.

6, method according to claim 5 is characterized in that, described auto adapted filtering step adopts the frequency domain adaptive filtering method.

7, method according to claim 6 is characterized in that, in the size of each filter coefficient update post detection filtering device coefficient, and when described filter coefficient is excessive, reduces its value.

8, method according to claim 7 is characterized in that, the value after described filter coefficient is lowered is: order $W(k+1)\&prime;=W(k+1)*\frac{\mathrm{Threshold}}{\left|\right|W_{\max }(k+1)\left|\right|},$

Described W (k+1)=[W ₀(k+1), W ₁(k+1) ... W _N-1(k+1)], be the coefficient of adaptive frequency domain filter, be the complex vector of a N dimension, N is a coefficient length;

Described W (k+1) ' is the value after described filter coefficient is lowered;

Described ‖ W _Max(k+1) ‖ is described W ₀(k+1), W ₁(k+1) ... W _N-1The maximal value of the mould that (k+1) each is plural;

Described $\mathrm{Threshold}=\sqrt{\frac{1+\cos \left(\mathrm{\&theta;}\right)}{1-\cos \left(\mathrm{\&theta;}\right)}},$ Wherein θ is the input angle maximal value of the voice signal of hope protection.

Images