CN106504763A - Multi-target Speech Enhancement Method Based on Microphone Array Based on Blind Source Separation and Spectral Subtraction - Google Patents

Abstract

The invention discloses a microphone array multi-target speech enhancement method based on blind source separation and spectral subtraction, which comprises the following steps: collecting multi-channel multi-target signals through a microphone array; performing band-pass filter processing on the collected single-channel signals respectively to obtain Shield non-speech segment noise and interference, and pre-emphasis processing; then perform speech windowing and framing processing to obtain frame signals, and use short-time Fourier transform to convert each frame to the frequency domain, and extract the amplitude spectrum and phase spectrum of each frame; Detect the start endpoint and end endpoint of the speech signal, estimate the noise power spectrum; reduce the background noise of the speech frame based on spectral subtraction; combine the output signal after spectral subtraction with the phase spectrum, and perform short-time Fourier inverse transform to obtain the speech signal in the time domain ; Finally, perform blind source separation to obtain each target signal. The implementation method of the invention is simple, has low resource requirements, low calculation complexity, and can realize multi-target signal enhancement.

Description

Multi-target Speech Enhancement Method Based on Microphone Array Based on Blind Source Separation and Spectral Subtraction

technical field

The invention belongs to the technical fields of signal processing technology and computer voice signal processing, and in particular relates to a voice enhancement method based on an array microphone.

Background technique

The goal of speech enhancement is to extract the original speech as pure as possible from the noisy speech signal, suppress background noise, improve the quality of speech and improve the comfort of the listener so that the listener will not feel tired. It is playing an increasingly important role in solving noise pollution, improving speech quality, and improving speech intelligibility. Speech enhancement technology is a problem that needs to be solved urgently after the speech signal processing develops to the practical stage. Anti-noise interference in speech recognition is an important link to improve the recognition rate. With the continuous expansion of speech recognition applications and entering the practical stage, there is an urgent need to adopt more effective speech enhancement technology to strengthen speech recognition features and make speech easy to recognize. Speech signal is a kind of complex nonlinear signal, how to separate the required speech signal from various mixed speech signals, especially from co-channel speech interference is a difficult digital signal processing problem, and any algorithm is impossible If the noise is completely filtered out, it is difficult to maintain high subjective and objective evaluation performance in the presence of all noise.

A typical workflow of the speech enhancement method based on the microphone array is shown in Figure 1, and the specific process mainly includes the following steps:

1) According to the requirements, design a microphone array structure that meets the requirements.

2) multi-channel voice signal acquisition system, used to collect multi-channel voice signals;

3) Perform preprocessing operations such as preprocessing, voice activation detection, channel delay estimation, and target signal orientation estimation on the collected multi-channel voice signals.

4) The array speech enhancement algorithm is used for speech enhancement to obtain a relatively pure speech signal.

In step 1), it is very important to design an appropriate microphone array structure.

Microphone array topology can be divided into one-dimensional linear array (including equidistant array, nested linear array and non-equidistant array), two-dimensional area array (including uniform and non-uniform circular array, square array) and three-dimensional stereo array. In practice, uniform line arrays, nested line arrays, and uniform area arrays are widely used. The research shows that the array topology has a great influence on the microphone array speech system. And the design of array topology is closely related to the choice of multi-channel signal model.

According to the distance of the sound source from the array, the sound signal model can be divided into a far-field model and a near-field model. The difference between the two is: the far-field model uses the plane wave model, which ignores the amplitude difference of the received signal of each channel, the source has only one incident angle relative to the array, and the delay length between each array element is linear; the near-field model Using a spherical waveform, it considers the amplitude difference between received signals, and there is an incident angle for each array element, and the delay length between each array element has no obvious relationship. There is no absolute standard for the division of near field and far field. It is generally believed that when the distance between the source and the center of the array is much greater than the signal wavelength, the source is in the far field; otherwise, it is in the near field.

Generally, a microphone array can be regarded as a device for spatial sampling. Similar to time sampling, the array sampling frequency must be high enough to avoid spatial ambiguity and avoid spatial aliasing. For an equidistant linear array, the spatial sampling rate is defined as: That is, the spatial sampling frequency U _s is determined by the microphone array spacing d. Considering that the difference between adjacent samples of the same signal is a phase shift, the normalized spatial frequency is defined as: Where λ represents the wavelength and Φ represents the angle of incidence. In order to avoid spatial aliasing, the normalized frequency U is required to satisfy: At this time, the corresponding range of incident angle is -90 ^° ≤ Φ ≤ 90 ^° , so the distance between adjacent microphones (microphone array distance) should be:

The above spatial sampling theorem reveals the relationship between the microphone array spacing, signal frequency and incoming wave direction (incident angle Φ). If the spatial sampling theorem is not satisfied, spatial aliasing occurs.

For a uniform linear microphone array, define r _m as the linear distance from the sound source to the center of the mth microphone array. Then the discrete signal output by the mth microphone can be expressed as: x _m [n]=s[n-Δn _m ]+η _m [n], where s[n] is the sound source signal, and Δn _m is the mth microphone The sample point delay between the received signal and the sound source signal, η _m [n] is the noise signal received by the mth microphone. Let Δτ _m be the time delay between the signal received by the mth microphone and the sound source signal, then the relationship is as follows: Among them, f _s is the time sampling frequency, and c is the propagation speed of the sound wave in space. From this, the array signal matrix output by the microphone array can be established:

x ₁ [n]=s[n-Δn ₁ ]+η ₁ [n]

x ₂ [n]=s[n-Δn ₂ ]+η ₂ [n]

x _N [n]=s[n-Δn _N ]+η _N [n]

N is the number of array elements of the array microphone.

In step 3), different enhancement methods can be used or subtracted.

In pre-processing, pre-emphasis and pre-filtering are determined by the characteristics of the speech signal. There are two purposes of pre-filtering: ① suppress all components whose frequency exceeds f _s /2 in each frequency domain component of the input signal to prevent aliasing interference; ② suppress 50Hz power frequency interference. Such a pre-filter must be a band-pass filter, and its upper and lower cut-off frequencies are respectively f _H and f _L , then f _H =3400Hz, f _L =60-100Hz, and sampling frequency f _s =16000Hz.

Since the average power spectrum of the speech signal is affected by glottal excitation and mouth and nose radiation, the high-frequency end drops at a rate of 6dB/octave above 800Hz. Therefore, when calculating the speech signal spectrum, the higher the frequency, the smaller the corresponding component, and the high-frequency Part of the spectrum is more difficult to find than the low frequency part, so pre-emphasis should be done in the pre-processing. The purpose of pre-emphasis is to enhance the high-frequency part, make the spectrum of the signal flat, and keep it in the entire frequency band from low frequency to high frequency. The same signal-to-noise ratio can be used to calculate the spectrum, so as to facilitate spectrum analysis or channel parameter analysis. Pre-emphasis can be realized by a pre-emphasis digital filter that improves high-frequency characteristics. It is generally a first-order digital filter. Based on its working principle, the corresponding emphasis method can be obtained as: s′(n)=s(n)-αs (n+1), in order to restore the original signal, it is necessary to de-emphasize the signal spectrum that has been pre-emphasized, that is, s''(n)=s'(n)+βs'(n+1) where, s( n) represents the sound source signal, s'(n) represents the signal after emphasis processing, and s"(n) represents the signal after de-emphasis processing, and β are aggravating factors, generally -0.8 to 0.95.

Since the speech signal is a non-stationary time-varying signal, its generation process is closely related to the movement of vocal organs. The state speed of the vocal organ is much slower than the speed of sound vibration, so the speech signal can be considered to be short-term stable. The study found that within the range of 5-50ms, the speech spectrum characteristics and some physical characteristic parameters basically remain unchanged. Therefore, the processing methods and theories in the stationary process can be introduced into the short-term processing of the speech signal, and the speech signal is divided into many short-term speech segments, and each short-term speech segment is called an analysis frame. In this way, processing each frame signal is equivalent to processing a continuous signal with fixed characteristics. Frames can be continuous or overlapped and divided into frames. Generally, the frame length is 10-30ms. When fetching data, the overlapping part of the previous frame and the next frame is called frame shift, and the ratio of frame shift to frame length is generally taken as 0~1/2. The extracted voice frame needs to be windowed, that is, the signal is multiplied by a certain window function w(n) to form a windowed voice. The main function of windowing is to reduce the spectrum leakage caused by framing processing. This is because framing is a sudden truncation of the speech signal, which is equivalent to the periodic convolution of the spectrum of the speech signal and the spectrum of the rectangular window function. Due to the high side lobe of the spectrum of the rectangular window, the spectrum of the signal will produce "tailing", that is, spectrum leakage. For this reason, the Hamming window can be used, because the Hamming window has the lowest side lobe, can effectively overcome the leakage phenomenon, has smoother low-pass characteristics, and obtains a smoother spectrum.

The estimation of the time delay between array elements plays a very important role in the entire microphone array speech enhancement algorithm: together with the signal frequency, it determines the directivity of the beam and is used to estimate the direction of the sound source. The accuracy of time delay estimation directly affects the performance of the speech processing system. Due to the spatial sampling of the speech signal by the microphone array, the signal received by the microphone has a certain delay relative to the reference microphone. In order to align the maximum direction of the beamformed output to the target signal source, maintaining the synchronization of the desired voice signals received by each microphone is an important means to solve this problem. Typical time delay estimation methods include generalized cross-correlation time delay estimation method, time delay estimation method based on adaptive filtering, adaptive eigendecomposition, high-order cumulant estimation method and so on. Among them, the generalized cross-correlation time delay estimation method is most commonly used. Assume that a pair of microphones receive the speech signal model as follows: x ₁ (t)=s(t)+η ₁ , x ₂ (t)=s(tD)+η ₂ , where s(t) is the sound source signal, x ₁ (t) and x ₂ (t) are the signals received by the two microphones respectively, D is the sound propagation delay between the two microphones, and η ₁ and η ₂ are additive background noises. Assuming that s(t), η ₁ , and η ₂ are not correlated with each other, signal amplitude attenuation is ignored here. Then the generalized cross-correlation function R ₁₂ (τ) between x ₁ (t) and x ₂ (t) is:

Where X ₁ (ω) and X ₂ (ω) are Fourier transforms of x ₁ (t) and x ₂ (t) respectively, and ψ ₁₂ (ω) is a generalized cross-correlation weighting function. Different weighting functions are selected for different situations, so that R ₁₂ (τ) has a relatively sharp peak, and the peak is the time delay between the two microphones.

Voice activation detection, also known as voice detection and voice endpoint detection, is used to accurately determine the starting point and end point of the input voice to ensure the good performance of the voice processing system. The processing methods for voice and noise are different. If it cannot be judged whether the current voice frame contains Noisy speech frames or noisy frames cannot be properly processed. In the speech enhancement system, in order to obtain more background noise characteristics, the speech endpoint detection focuses more on how to accurately detect the silent segment. Both the learning of phonetic knowledge and the accumulation of noise source information estimation rely on accurate endpoint detection. Common voice activation detection is performed based on voice frames, and the length of the voice frames ranges from 10 to 30 ms. The method of voice activation detection can be summarized as: extracting one or a series of contrast feature parameters from the input signal, and then comparing it with one or a series of threshold thresholds. If it exceeds the threshold, it means that there is currently a sound segment, otherwise it means that there is currently no sound segment.

Speech detection generally has two steps:

The first step: based on the characteristics of the speech signal. Use parameters such as energy, zero-crossing rate, entropy, pitch, and their derivative parameters to judge speech/non-speech segments in the signal stream.

Step 2: After the voice signal is detected in the signal stream, it is judged whether this is the start point or the end point of the voice. In the speech system, it is easier to have mid-sentence pauses (non-speech) due to the variable background of the signal and natural dialogue patterns, especially the silent gap before the burst initial. Therefore, this judgment of beginning or end is particularly important.

At present, the methods adopted for voice endpoint detection can be roughly divided into two categories:

The first type is the method of speech signal endpoint detection based on HMM model in noisy environment, which requires the background noise to be stable and the signal-to-noise ratio is high.

The second type of method is based on the short-term energy detection algorithm of the signal. It determines the energy threshold through the statistics of the background noise energy, and uses the energy threshold to determine the starting point of the speech signal.

In step 4), a relatively pure speech signal is obtained by using a speech enhancement algorithm.

Speech enhancement technology can be mainly divided into methods based on single channel and methods based on multi-channel array microphones. There are many kinds of single-channel speech enhancement methods, most of which are based on various noise cancellation methods combined with the characteristics of the speech signal to study targeted algorithms. The theory is mature and the simplest and most effective is spectral subtraction (SS: Spectral Subtraction) speech enhancement. The sound pickup by a single sensor will be limited by the venue, distance, and application occasions, so the sound pickup effect will be greatly reduced, and subsequent speech enhancement will be difficult.

The basic principle of spectral subtraction is: in the frequency domain, the power spectrum of the noisy speech is subtracted from the power spectrum of the noise to obtain the power spectrum estimate of the speech. Recover the time domain signal. Considering that the human ear is not sensitive to the phase, the phase used in the phase recovery is the phase information of the noisy speech. Since speech is short-term stationary, it is considered to be a stationary random signal in short-term spectrum amplitude estimation.

Suppose s(n), η(n) and x(n) represent speech, noise and noisy speech respectively, and S(ω), Γ(ω) and X(ω) represent their short-term spectrum respectively. Assume that s(n), η(n) are uncorrelated and the noise is additive noise. Then the additive model of the signal is obtained: x(n)=s(n)+η(n), and the signals after windowing processing are respectively expressed as x _w (n), s _w (n), η _w (n) , then: x _w (n) = s _w (n) + η _w (n), do Fourier transform on it, get: X _w (ω) = S _w (ω) + Γ _w (ω), so for The power spectrum is: |X _w (ω)| ² ＝|S _w (ω)| ² + |X _w ()| ² is estimated from observed data, and the remaining terms must be approximated by statistical means. Since s(n) and η(n) are independent, the statistical mean value of the cross power is 0, so the estimate of the original speech is: Among them, the estimated value It cannot be guaranteed to be non-negative, because there is an error in estimating the noise. When the estimated noise average power is greater than the noisy speech power of a certain frame, the estimated value obtained by the frame There will be negative cases. These negative values can be changed to positive values by changing their signs, or they can be directly set to zero. Will The time domain estimation of the speech signal can be obtained by recovering the phase and doing the inverse short-time Fourier transform IFFT:

Currently, microphone array speech enhancement algorithms mainly include beamforming, subspace decomposition, and blind source separation. Among them, Blind Source Separation (BSS) refers to the process of recovering the source signal only from the observed signal when the source signal and the mixing method are not known or cannot be obtained. That is, blind source separation does not depend on the prior conditions of the current event, and can perform speech enhancement with fewer microphones. The central problem of this algorithm is to separate the speech of each speaker in the case of mutual interference and aliasing of multiple people's speech , to achieve the purpose of enhancing the voice of each target.

Independent component analysis (ICA) is one of the effective methods for blind signal separation, which belongs to linear instantaneous mixed blind signal processing. This method does not depend on the detailed knowledge of the source signal type or the precise identification of the signal transmission system characteristics. Redundancy Cancellation Technology. According to different cost functions, this method can obtain different ICA algorithms, such as information maximization (infomax) algorithm, Fast ICA algorithm, maximum entropy (ME) and minimum mutual information (MM I) algorithm, maximum likelihood (ML) algorithm etc. The basic principle is: the obtained signal is regarded as the target signal mixed through a linear transformation. In order to obtain the target signal, it is necessary to find an inverse linear transformation to decompose the obtained signal, so as to achieve the purpose of source separation.

In the case of no noise, use X=[x ₁ (t) x ₂ (t) ... x _N (t)]' to represent a set of observation signals received by the microphone array, where t is the time or sample number, N is the number of microphones, assuming that it is linearly mixed from independent components, that is, where A is an unknown full-rank matrix. So the vector expression of its signal model is X=AS.

In the presence of noise, it is assumed that the noise is additive. Then the expression of its signal model is: X=AS+Γ, where Γ=[η ₁ η ₂ ... η _N ] is the noise vector. Transform X=AS+Γ to get: X=A(S+Γ ₀ ), Γ = AΓ ₀ , Therefore, it can be concluded that the noisy signal model It is still the basic ICA model, but the independent components are transformed from S to Under the basic signal model of ICA, assuming that the separation matrix to be obtained is W, and the signal matrix after separation is Y, the following expression is given: Y=WX=WAS. The ultimate goal of ICA is to find an optimal or better separation matrix W so that each signal in the separated signal matrix Y is independent of each other and as close as possible to the source signal.

At present, speech enhancement based on a microphone array is performed on a single target, which limits the effective sound pickup effect of the array pickup device, and the traditional single-target enhancement cannot meet the needs of practical applications.

Contents of the invention

In order to solve the current technical problem of multi-target enhancement based on array speech signals, the present invention proposes a microphone array multi-target speech enhancement method based on blind source separation and spectral subtraction.

The microphone array multi-target voice enhancement method based on blind source separation and spectral subtraction of the present invention comprises the following steps:

Step 1: collect the noisy speech signal through the microphone array of the two-dimensional area array, and obtain the acquisition signals of each channel of the microphone array, wherein the number of microphone arrays is greater than or equal to 4;

Step 2: Execute steps 201-205 on the collected signals of each channel respectively:

Step 201: Perform band-pass filtering processing on the collected signal to shield non-speech segment noise and interference; then perform pre-emphasis processing on the band-pass filtered signal, divide into frames, and add window processing to obtain a frame signal;

Then perform frequency domain conversion on each frame signal, that is, perform short-time Fourier transform on each frame signal, and calculate the power spectrum of each frame; at the same time, calculate and retain the phase spectrum of each frame for phase recovery in the process of spectrum subtraction;

Step 203: Perform speech detection on the frame signal of each frame, determine whether the current frame is a speech frame or a noise frame, and estimate the noise power spectrum based on the noise frame;

Step 204: remove the noise power spectrum in the power spectrum of the speech frame based on spectral subtraction, and obtain the speech power spectrum estimation of each frame;

Step 205: Estimate the square root of the speech power spectrum, perform phase recovery based on the phase spectrum of the corresponding frame, and then perform inverse short-time Fourier transform to obtain the time domain estimation signal of the speech frame;

In step 2, signal preprocessing is performed on the single-channel acquisition signal, and the acquisition signal is divided into many short-term speech segments (with noise), that is, frame signals, and then spectral subtraction is performed on the frame signals of each frame to reduce the Background noise for speech frames.

Step 3: Carry out source separation processing to the time-domain estimation signals of the speech frames of all channels using the blind source separation method to obtain target signals of different sources;

Step 4: De-emphasis, de-windowing, and frame reorganization are performed on the target signal of the same source to obtain target speech signals of different sources.

In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: (1) solve the traditional single-channel speech enhancement method to process environmental background noise, the algorithm is simple, and the technical problems of low demand for resources; (2) ) no longer rely on the array signal processing algorithm for spatial filtering, and do not need to consider the broadband beam algorithm, which reduces the complexity of the algorithm structure; (3) uses the blind source separation algorithm to realize the enhancement of the target signal, no longer single or rotating Single target signal enhancement.

Description of drawings

FIG. 1 is a schematic diagram of a traditional speech enhancement system.

Fig. 2 is a schematic diagram of an implementation system of a specific embodiment of the present invention.

Fig. 3 is a flowchart of speech detection.

Fig. 4 is a flow chart of a method for spectral subtraction single-channel speech enhancement.

detailed description

In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the implementation methods and accompanying drawings.

Referring to Fig. 2, the multi-target speech enhancement method of the present invention, at first carries out signal preprocessing to each single-channel signal (speech signal) that the microphone array of two-dimensional area array gathers, and the speech signal of single-channel is divided into a lot of short-term speech segment to obtain frame signals for subsequent speech activation detection and spectral subtraction processing. The signal preprocessing includes band-pass filtering, pre-emphasis processing, overlapping framing, and Hamming window processing.

Perform speech activation detection and spectral subtraction processing on the single-channel frame signals, and then perform blind source separation on all channels of the same speech frame to obtain target signals of different source signals, and then correspond to the inverse operation of signal preprocessing to perform the same source signal The target signal is de-emphasized, the Hamming window is removed, and the frame is reorganized to obtain each target speech signal, so as to realize the enhancement processing of multi-target speech.

According to the characteristics of the sound source and noise field in the indoor environment, the multi-channel noisy speech signal in the actual environment is modeled by using the diffuse noise field model and the near-field sound source model. Speech signals in the space are collected through an 8×8 planar array composed of 64 microphones.

Use X=[x ₁ (t) x ₂ (t) ... x _j (t) ... x _N (t)]' to represent the noisy speech signal output by each channel, where j represents the serial number of the microphone channel.

Then the array signal (frame signal) obtained after signal preprocessing is carried out to the noisy speech signal output by each channel is X _pw , then X _pw =[x _1pw (n) x _2pw (n) ... _jpw (n) ... x _Npw (n)]', where n=1,2,...L, L is the frame length, w is the frame number.

Make short-time Fourier transform on the frame signal X _pw to obtain the amplitude spectrum |X _pw (ω) and the phase spectrum Φ _pw (ω). Among them, ω is the frequency sampling point, which is the uniform sampling of N equal parts of the angular frequency from 0 to 2π. So have:

|X _pw |＝[|X _1pw (ω)| |X _2pw (ω)| … |X _jpw (ω)| … |X _Npw (ω)|]′

Utilize |X _pw |=[|X _1pw (ω)| |X _2pw (ω)| ... |X _jpw (ω)| ...|X _Npw (ω)|] ', carry out speech starting according to the flow chart shown in Figure 3 The detection of the start endpoint and the end endpoint is to judge whether the current frame is a noise frame or a voice frame, and use the judgment result to perform spectrum reduction and noise reduction. Wherein, the specific process of the detection of speech start endpoint (start frame) and end endpoint (end frame) is:

use the formula Calculate the voice energy of each frame, where N is the frame length, w is the number of the frame, 1≤w≤L, L is the number of frames, and ω is each point in each frame;

Initialize the threshold T, and set the initial value of the threshold T through the statistics of background noise energy

Then classify each frame based on the threshold T, determine whether the current frame is a noise frame or a speech frame, and update the threshold T based on the noise frame of the latest k frames:

a. Calculate the voice energy M _w of the current frame, if M _w is greater than T, then determine that the current frame is a voice frame, otherwise it is determined to be a noise frame;

b. If the current frame is a noise frame, the threshold T is updated based on the nearest k (experience value, usually greater than or equal to 10) frame noise frame:

b1: Calculate the average speech energy EMN, speech energy maximum value and energy minimum value EAX, EMIN of the noise frame of the latest k frames;

b2: According to the formula T=min[a×(EAX-EMIN)+EMN,b×EMN](0<a<1,1<b<10) to obtain the updated threshold T;

c. If the current frame is a speech frame, it is determined whether all frames have been processed, and if so, the endpoint detection is completed; otherwise, continue to repeat steps a to c for the next frame.

Furthermore, the short-term zero-crossing rate can also be used to check the judgment results of the speech frame and the noise frame, so as to prevent misjudgment.

Referring to Figure 4, based on all detected noise frames, the noise power spectrum can be estimated, and then the estimated noise of each speech frame is removed based on spectral subtraction, that is, the power spectrum of the speech frame is subtracted from the currently estimated noise power spectrum to obtain the speech The speech power spectrum of each frame is estimated, and then the square root of the speech power spectrum is estimated, and the phase is restored based on the phase spectrum of each speech frame, and then the short-time inverse Fourier transform is performed to obtain the time domain estimation signal of the speech frame, that is, the single-channel Enhance speech signal.

After the above is completed, the natural gradient ICA is used to complete the blind source separation, and the specific processing process is as follows:

(1) If the mean value of the current observed signal X (multiple single-channel enhanced speech signal sequences) is not zero, then first subtract its mean value from the observed signal X;

(2) select a matrix B, make covariance matrix E{VV ^T } be identity matrix I, wherein V=BX, be irrelevant between each component of vector V, and have unit variance;

(3) Whitening processing based on singular value decomposition: first estimate the variance R _x = E{XX ^T } of X, and R _x is a real Hermitian matrix; secondly, perform singular value decomposition on R _x The column of U=[u ₁ , u ₂ ,..., u _n ] is the left singular vector of R _x , and the purpose of whitening processing is to weaken the correlation of speech signals after mixing;

(4) Since σ ₁ ≥σ ₂ ≥...≥σ _m >0; σ _m+1 =...σ _n ; (m≤n), the number of source signals is estimated to be m;

(5) Finally, perform an orthogonal transformation:

U＝[u ₁ , ₂ ,…, _n ]

=BX, U _m =[u ₁ , ₂ ,..., _m ]

According to the formula Get a reply signal Y, where, P can be called both the performance matrix and the convergence matrix, W represents the separation matrix in ICA, A represents the mixing matrix in ICA, and S represents the source signal.

The above is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless specifically stated, can be replaced by other equivalent or alternative features with similar purposes; all the disclosed features, or All method or process steps may be combined in any way, except for mutually exclusive features and/or steps.

Claims (3)

1. based on blind source separating and the microphone array multiple target sound enhancement method of spectrum-subtraction, it is characterised in that including following Step：

Step 1：Noisy Speech Signal gathered by the microphone array of two-dimensional array, each passage for obtaining microphone array is adopted Collection signal, wherein microphone array column number are more than or equal to 4；

Step 2：Collection signal execution step 201～205 to each passage respectively：

Step 201：Bandpass filtering treatment, shielding non-speech segment noise and interference are carried out to gathering signal；Again to bandpass filtering after Signal carry out preemphasis process, framing, windowing process obtain frame signal；

Short time discrete Fourier transform is carried out to the frame signal of every frame, and calculates the power spectrum and phase spectrum of each frame；

Step 203：Speech detection is carried out to the frame signal of every frame, judges that present frame is speech frame or noise frame, based on noise Frame estimating noise power is composed；

Step 204：Noise power spectrum in the power spectrum of speech frame is removed based on spectrum-subtraction, the phonetic speech power spectrum for obtaining every frame is estimated Meter；

Step 205：To phonetic speech power Power estimation evolution, and after the phase spectrum based on corresponding frame carries out phase recovery, then carry out short When inverse fourier transform, obtain speech frame time domain estimate signal；

Step 3：Signal carries out information source separating treatment using blind source separating method to be estimated to the time domain of the speech frame of all passages, is obtained The echo signal of various information source；

Step 4：The echo signal of same information source is carried out postemphasising, goes window, frame reorganization, the target language of various information source is obtained Message number.

2. the method for claim 1, it is characterised in that the microphone array is classified as 8*8 rectangle plane arrays, each battle array Unit is uniformly distributed on the whole, can carry out Subarray partition, and different submatrixs can work independently.

3. method as claimed in claim 1 or 2, it is characterised in that in step 3, using self adaptation natural gradient blind source separating Carry out information source separating treatment.