Wiener gain design method

Abstract

The invention discloses a Wiener gain design method which comprises the following steps: Wiener gain estimation is carried out, the Wiener gain obtained by estimation is used as a post-filter, and the Wiener gain estimation process comprises the following steps: signal-to-noise ratio estimation is carried out according to an output signal output by a beam former to obtain a signal-to-noise ratio estimation value; in the noise and reverberation environment, performing coherent scattering ratio estimation according to the signal-to-noise ratio estimation value and the observation signal acquired by the microphone array to obtain a coherent scattering ratio estimation value; and performing gain estimation according to the signal-to-noise ratio estimation value and the coherent scattering ratio estimation value to obtain a Wiener gain. According to the method, the Wiener gain is designed by comprehensively considering the CDR and the SNR, and the performance of the speech enhancement technology in the aspects of speech noise reduction, distortion control, reverberation removal capability and the like is greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of speech enhancement, and in particular to a design method of Wiener gain. Background Art

[0002] With the popularity of smart devices and the rise of remote working modes, the demand for remote voice communication and human-computer interaction systems is growing. However, in these systems, since the sound source signal will inevitably be reflected by hard surfaces such as walls, floors or ceilings during the propagation process, the received observation signal contains a large number of late reflections, namely reverberation components. In addition, the sound source signal is also disturbed by background noise. These reverberation components and background noise seriously reduce the quality and intelligibility of speech, and have a significant impact on the performance of signal processing algorithms such as speech recognition and sound source localization, limiting the improvement of user experience.

[0003] The Wiener beamformer is a spatiotemporal filtering technique widely used in the field of acoustic signal processing to extract the target sound source signal under the interference of noise and reverberation. This Wiener beamformer can be decomposed into a combination of a minimum variance distortionless response (MVDR) beamformer and a Wiener gain. Therefore, in order to ensure the best performance of the Wiener beamformer, the MVDR beamformer and the Wiener gain must be carefully designed.

[0004] A large amount of literature research focuses on the design of MVDR beamformers, aiming to achieve high signal-to-noise ratio (SNR) gain while ensuring its robustness in practical applications. To this end, various techniques have been proposed, including diagonal-loading MVDR beamformers, worst-case MVDR beamformers, and simplified MVDR beam formers, where the noise field is modeled as spatial white noise or scattered noise.

[0005] Similarly, the design of the optimal Wiener gain has also received extensive attention, playing an important role in both single-channel and multi-channel speech enhancement. However, the traditional Wiener gain formula is usually based on the signal-to-noise ratio (SNR) or the coherence diffusion ratio (CDR). Among them, the method based on the signal-to-noise ratio (SNR) is mainly committed to suppressing background additive noise. The key to this method is to accurately estimate the signal-to-noise ratio, because a more accurate SNR estimate can effectively reduce the distortion in the speech enhancement process. In contrast, the method based on the coherence diffusion ratio (CDR) takes reverberation suppression as the core goal. The key is to obtain an accurate estimate of the CDR. In order to estimate the CDR, the late reverberation signal is usually assumed to be incoherent diffuse noise. Specifically, it is necessary to calculate the coherence coefficient corresponding to the distance between the sensors and the frequency of interest. By using the covariance matrix model of the array observations and the prior information of the noise coherence matrix, a CDR estimator can be derived. Initially, the CDR estimator was developed for a binary array and later extended to arrays containing more than two sensors.

[0006] However, existing CDR estimators are unable to consider noise and reverberation simultaneously, which makes them work poorly in practical environments where background noise and reverberation coexist. Summary of the invention

[0007] The object of the present invention is to provide a method for designing Wiener gain by comprehensively considering CDR and SNR.

[0008] To achieve the above object, the present invention proposes a design method of Wiener gain, comprising:

[0009] S1, in an environment where noise and reverberation coexist, processing an observation signal collected by a microphone array through a beamformer to output a beam output signal, wherein the observation signal includes a source signal and a noise source, and the noise source includes at least an additive noise signal and a reverberation signal;

[0010] S2, performing Wiener gain estimation, and using the estimated Wiener gain as a post filter, the Wiener gain estimation process includes:

[0011] S21, performing signal-to-noise ratio estimation according to the output signal in step S1 to obtain a signal-to-noise ratio estimation value;

[0012] S22, performing coherent scatter ratio estimation according to the signal-to-noise ratio estimation value and the observation signal collected in step S1 to obtain a coherent scatter ratio estimation value;

[0013] S23, performing gain estimation according to the estimated value of the signal-to-noise ratio and the estimated value of the coherent scattering ratio to obtain the Wiener gain;

[0014] S3, processing the output signal through a post-filter to output a final source signal.

[0015] In a preferred embodiment, in step S1, the observation signal is represented as:

[0016] y(n,k)=[Y1(n,k)Y2(n,k)···Y M (n,k)] T =d(k)S(n,k)+r(n,k)+v(n,k);

[0017] Among them, y(n,k) is the observed signal vector, Y m (n,k) is the observation signal collected by the mth microphone, m = 1, 2, ···M, M is the number of elements in the microphone array, [·] T represents the transpose operation of a vector or matrix, d(k) represents the steering vector of the array, S(n,k) represents the source signal, r(n,k) represents the reverberation signal vector in the observed signal, v(n,k) represents the additive noise signal vector in the observed signal, n is the frame index, k=2πf represents the angular frequency, and f represents the frequency;

[0018] The output signal is expressed as:

[0019] z(n,k)=h H (k)y(n,k);

[0020] Wherein, z(n,k) is the output signal vector, h(k) represents the weight vector of the beamformer, [·] H represents transposed conjugate.

[0021] In a preferred embodiment, in S21, a direct decision method is used to estimate the signal-to-noise ratio.

[0022] In a preferred embodiment, the calculation formula of the signal-to-noise ratio estimation value is:

[0023] SNR(n,k)=αSNR(n-1,k)+(1-α)·max(ξ(n,k)-1,0);

[0024] Wherein, α∈(0,1) is a hyperparameter, n is a frame index, k=2πf represents the angular frequency, f represents the frequency, ζ represents the posterior signal-to-noise ratio, and ζ(n,k) is expressed as:

[0025]

[0026] Among them, z(n,k) is the output signal vector, φ V (n,k) is the variance of the additive noise signal.

[0027] In a preferred embodiment, in S22, the calculation process of the coherent scatter ratio estimation value is:

[0028] Calculate the estimated value of the coherent scattering ratio of the microphone pair according to the pseudo-coherence matrix and the estimated value of the signal-to-noise ratio of the observation signal, the additive noise signal and the reverberation signal collected by the microphone pair of any two channels in the microphone array;

[0029] The coherent scatter ratio estimates of all microphone pairs in the microphone array are then averaged.

[0030] In a preferred embodiment, the calculation formula of the coherent scatter ratio estimation value is:

[0031]

[0032] Where M is the number of elements in the microphone array, CDR i,j (k) is the estimated value of the coherent scatter ratio of the microphone pair, i = 1, 2 ... M-1, j = 0, 1 ... i-1, k = 2πf represents the angular frequency, and f represents the frequency;

[0033] The coherent scatter ratio estimate for each microphone pair is given by:

[0034]

[0035] in,

[0036]

[0037] [Γ Y (k)] i,j is the pseudo-coherence matrix of the observation signal collected by the microphone pair, [Γ V (k)] i,j is the pseudo-coherence matrix of the additive noise signal collected by the microphone pair, [ Γ R(k)] i,j is the pseudo-coherence matrix of the reverberation signal collected by the microphone pair, represents the real part operation, [·] * represents the conjugate transpose, and SNR(k) is the signal-to-noise ratio estimate.

[0038] In a preferred embodiment, in S23, the calculation process of the Wiener gain includes:

[0039] Constructing a Wiener gain function based on a coherent spread ratio and a signal-to-noise ratio from the output signal of the beam in step S1;

[0040] The signal-to-noise ratio estimation value obtained in step S21 and the coherent scatter ratio estimation value obtained in step S22 are substituted into the Wiener gain function to calculate the Wiener gain.

[0041] In a preferred embodiment, in S23, the Wiener gain is expressed as:

[0042]

[0043] in,

[0044]

[0045] h(k) represents the weight vector of the beamformer, [·] H represents the transposed conjugate, Γ R (k),Γ V (k) represents the pseudo-coherence matrix of the reverberation signal vector and the additive noise signal vector, respectively, CDR(n,k) represents the estimated value of the coherent scattering ratio, and SNR(n,k) represents the estimated value of the signal-to-noise ratio.

[0046] In a preferred embodiment, the S2 further includes: S24, adding two hyperparameters for balancing the attenuation degree of noise and reverberation to the Wiener gain, and after adding the hyperparameters, the Wiener gain is expressed as:

[0047]

[0048] Among them, β1≥0, β2≥0 are two hyperparameters.

[0049] In a preferred embodiment, in S3, the final source signal is expressed as:

[0050]

[0051] in, is the final source signal, G(n,k) is the Wiener gain, Z(n,k) is the output signal of the beam, n is the frame index, k=2πf represents the angular frequency, and f represents the frequency.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. Aiming at the coexistence problem of noise and reverberation, the present invention proposes a novel beamformer design method, which adopts the classic beamforming + post-filtering framework and comprehensively considers CDR and SNR to design the Wiener gain, thereby greatly improving the performance of speech enhancement technology in speech noise reduction, distortion control, and dereverberation capability.

[0054] 2. The present invention also uses two hyperparameters to control the attenuation degree of noise and reverberation, so as to flexibly adjust the balance between noise reduction and reverberation suppression to improve the signal-to-noise ratio and reduce noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the flow of the design method of Wiener gain of the present invention;

[0056] Figure 2 It is a signal flow diagram of the method of the present invention;

[0057] Figure 3a A schematic diagram of the trade-off between DRR and SNR Gain under the conditions of β1, β2 as a function of the method of the present invention and other methods;

[0058] Figure 3b A schematic diagram of the trade-off between LSD and SNR Gain under the conditions of β1, β2 under the conditions of the method of the present invention and other methods;

[0059] Figures 4a to 4c Schematic diagram of SNR Gain, LSD and DRR as functions of input signal-to-noise ratio for the method of the present invention and other methods;

[0060] Figure 5a to 5c Schematic diagram of SNR Gain, LSD and DRR as a function of reverberation time for the method of the present invention and other methods;

[0061] Figure 6 Schematic diagram of SRMR as a function of input signal-to-noise ratio for the method of the present invention and other methods;

[0062] Figure 7 The logarithmic kurtosis ratio of the method of the present invention and other methods is G min Function diagram. DETAILED DESCRIPTION

[0063] The specific embodiments of the present invention are described in detail below, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0064] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0065] The present invention aims at the coexistence problem of noise and reverberation, and proposes a design method of Wiener gain. The method adopts the classic beamforming + post-filtering framework, comprehensively considers the coherent scattered signal energy ratio (CDR) and signal-to-noise ratio (SNR) to design the Wiener gain, and then uses the designed Wiener gain as the post-filtering gain, and applies it to the signal processing process after beamforming. Through these processing steps, a significantly enhanced speech signal is obtained. In addition, the present invention also introduces two hyperparameters to control the attenuation degree of background noise and reverberation, so as to flexibly adjust the balance between noise reduction and reverberation suppression, so as to improve the signal-to-noise ratio and reduce noise interference.

[0066] Combination Figure 1 and Figure 2 As shown, a design method of Wiener gain disclosed in the present invention comprises the following steps:

[0067] S1, in an environment where noise and reverberation coexist, the observation signal collected by the microphone array is processed by a beamformer to output a beam output signal, wherein the observation signal includes a source signal and a noise source, and the noise source includes at least an additive noise signal and a reverberation signal.

[0068] Specifically, for the actual environment where noise and reverberation coexist, reverberation and additive noise are uniformly regarded as noise sources that need to be suppressed. The present invention uses a microphone array composed of M (m = 1, 2, ..., M) array elements to collect observation signals, and the signal observed by the mth microphone is defined as:

[0069] y m (t) = g m (t)*s(t)+v m (t)

[0070] =g m,early (t)*s(t)+g m,late (t)*s(t)+v m (t);

[0071] Where t is the discrete sampling time point, * represents the linear convolution operation, s(t), v m (t) represent the source signal and additive noise signal respectively, g m (t) represents the impulse response from the sound source to the mth microphone, g m,early (t) and g m,late (t) represent the early reflection and late reverberation respectively, which satisfy g m (t) = g m,early (t)+g m,late (t).

[0072] After short-time Fourier transform, the above formula can be re-expressed as:

[0073] Y m (n,k)=X m (n,k)+R m (n,k)+V m (n,k);

[0074] Among them, Y m (n,k),X m (n,k), R m (n,k), V m (n,k) represent y in the above formula respectively m (t), g m,early (t)*s(t),g m,late (t)*s(t),v m (t) is the short-time Fourier transform of n. n represents the frame index, k = 2πf represents the angular frequency, and f represents the frequency.

[0075] Put all the above observation signals into a vector, and the observation signal vector collected by the microphone array is expressed as:

[0076] y(n,k)=[Y1(n,k) Y2(n,k) … Y M (n,k)] T

[0077] =d(k)S(n,k)+r(n,k)+v(n,k);

[0078] Among them, y(n,k) is the observed signal vector, Y m (n,k) is the observation signal collected by the mth microphone, [·] T Represents the transpose operation of a vector or matrix;

[0079] r(n,k)=[R1(n,k) R2(n,k) … R M(n,k) ] T ;

[0080] v(n,k)=[V1(n,k) V2(n,k) … V M (n,k)] T ;

[0081] S(n,k), r(n,k), v(n,k) represent the source signal vector, reverberation signal vector and additive noise signal vector in the observed signal respectively, and d(k) represents the steering vector of the array.

[0082] Assuming that the source signal, reverberation signal and additive noise signal are uncorrelated, the variance of the beam output signal can be expressed as:

[0083] φ Z (n,k)=E[|Z(n,k)|2

[0084] =|h H (k)d(k)| 2 φ S (n,k)+h H (k)Φ R (n,k)h(k)+h H (k)Φ V (n,k)h(k);

[0085] Among them, E[·] represents the mathematical expectation, φ S (n, k) represents the variance of the source signal that needs to be retained, Φ R (n,k) and φ V (n, k) represent the covariance matrix of the reverberation signal vector and the additive noise signal vector, respectively, which can be expressed as:

[0086] Φ R (n,k)=Φ R (k)φ R (n,k);

[0087] Φ V (n,k)=Γ V (k)φ V (n,k);

[0088] Among them, Γ R (k),Γ V (k) represent the pseudo-coherence matrices of the reverberation signal vector and the additive noise signal vector, respectively. The trace tr[Γ R (k)] = tr[Γ V (k)]=M;φ R (n,k),φ V (n, k) represent the variance of the reverberation signal and the additive noise signal respectively.

[0089] At this time, the present invention applies a beam h(k) with a length of M to the above-mentioned observation signal vector y(n, k) of the array, and the output signal of the beam is obtained as follows:

[0090] Z(n,k)=h H (k)y(n,k)

[0091] =h H (k)d(k)S(n,k)+h H (k)r(n,k)+h H (k)v(n,k);

[0092] Among them, h H (k)r(n,k) and h H(k)v(n,k) are the residual reverberation signal and residual additive noise signal in the array output signal, respectively. H (k)d(k)S(n,k) represents the source signal that needs to be retained.

[0093] The present invention adopts a robust super-directional beamformer as the desired beamformer h(k) to enhance the signal reception in the target direction while suppressing noise interference in other directions, which is specifically expressed as:

[0094]

[0095] Where h(k) represents the weight vector of the beamformer, i.e., the beam h(k) of length M, k = 2πf represents the angular frequency, f represents the frequency, [·] H represents the transposed conjugate, Γ R,∈ (k) = Γ R (k)+∈I,Γ R (k) is the pseudo-coherence matrix of the reverberation signal, and its calculation formula is: In the formula, c = 340m / s is the speed of sound in air, Δ i,j is the distance between the i-th and j-th microphones in the microphone array, i=1,2……M-1,j=0,1……i-1.∈=10 -3 , I represents the unit matrix. d(k) is the array phase delay vector corresponding to the desired signal direction, that is, the above-mentioned steering vector, which can be expressed as:

[0096]

[0097] Where τ0 = δ / c, δ is the adjacent spacing of the microphone array, and θ0 is the direction of the source signal.

[0098] S2, perform Wiener gain estimation, and use the estimated Wiener gain as a post filter.

[0099] Specifically, the Wiener gain estimation process includes:

[0100] S21, performing signal-to-noise ratio estimation according to the output signal in step S1 to obtain a signal-to-noise ratio estimation value.

[0101] Specifically, the present invention adopts a decision-directed method to estimate the signal-to-noise ratio, and the calculation formula is specifically expressed as:

[0102] SNR(n,k)=αSNR(n-1,k)+(1-α)·max(ζ(n,k)-1,0);

[0103] Wherein, α∈(0,1) is a hyperparameter, n is a frame index, k=2πf represents the angular frequency, f represents the frequency, ζ represents the posterior signal-to-noise ratio, and ζ(n,k) is expressed as:

[0104]

[0105] Among them, Z(n,k) is the above output signal vector, φ V (n,k) is the variance of the additive noise signal.

[0106] S22, performing coherent scatter ratio estimation according to the signal-to-noise ratio estimation value and the observation signal collected in step S1 to obtain a coherent scatter ratio estimation value.

[0107] Specifically, in this step, the calculation process of the coherent scatter ratio estimation value is as follows:

[0108] Calculate the estimated value of the coherent scattering ratio of the microphone pair according to the pseudo-coherence matrix and the estimated value of the signal-to-noise ratio of the observation signal, the additive noise signal and the reverberation signal collected by the microphone pair of any two channels in the microphone array;

[0109] Specifically, the covariance matrix of the observed signal can be expressed as:

[0110] φ Y (n,k)=φ S (n,k)dd H +φ R (n,k)Γ R (k)+φ V (n,k)Γ v (k) = φ S (n,k)Γ S (k)+φ R (n,k)Γ R (k)+φ V (n,k)Γ v (k);

[0111] Among them, Φ Y (n, k) is the covariance matrix of the observed signal, which can be obtained by recursive averaging:

[0112] Φ Y (n,k)=λΦ Y (n-1,k)+(1-λ)y(n,k)y H (n,k);

[0113] Among them, λ∈(0,1) is the forgetting factor.

[0114] φ S (n,k) is the variance of the source signal, Γ S (k) = ddH Represents the pseudo-coherence matrix of the source signal. R (n,k),φ V (n, k) represent the variance of the reverberation signal and the additive noise signal, respectively, Γ R (k),Γ V (k) represents the pseudo-coherence matrix of the reverberation signal and the additive noise signal respectively. Under the assumption that the source signal, late reverberation and additive noise signal are uncorrelated, we have φ Y (n,k)=φ S (n,k)+φ R (n,k)+φ V (n,k).

[0115] The pseudo-coherence matrix of the observed signal can be calculated according to the following formula:

[0116]

[0117] Substituting the covariance matrix of the observed signal into the above formula, the coherence matrix of the source signal can be further expressed as:

[0118]

[0119] By imposing an absolute value constraint on the coherence coefficient of the source signal, the CDR can be estimated from the observed signals of any two channels (i.e., microphone pairs). If the SNR is given, the formula for the coherent scatter ratio estimate of each microphone pair is:

[0120]

[0121] in,

[0122]

[0123] [Γ Y (k)] i,j is the pseudo-coherence matrix of the observation signal collected by the microphone pair, [Γ V (k)] i,j is the pseudo-coherence matrix of the additive noise signal collected by the microphone pair, [Γ R (k)] i,j is the pseudo-coherence matrix of the reverberation signal collected by the microphone pair, represents the real part operation, [·] * represents the conjugate transpose, and SNR(k) is the signal-to-noise ratio estimate.

[0124] The coherent scatter ratio estimates of all microphone pairs in the microphone array are then averaged.

[0125] Specifically, the calculation formula of the coherent scatter ratio estimate is:

[0126]

[0127] Among them, CDR i,j (k) is the estimated value of the coherent scattering ratio of the above microphone pair, i = 1, 2...M-1, j = 0, 1...i-1, k = 2πf represents the angular frequency, and f represents the frequency.

[0128] S23, performing gain estimation according to the estimated value of the signal-to-noise ratio and the estimated value of the coherent scattering ratio to obtain the Wiener gain.

[0129] Specifically, in this step, the calculation process of the Wiener gain specifically includes:

[0130] A Wiener gain function based on a coherent spread ratio and a signal-to-noise ratio is constructed from the output signal of the beam in step S1.

[0131] Specifically, in step S1, by considering the covariance matrix of the reverberation signal vector and the additive noise signal vector, the variance of the output signal Z(n, k) of the beam can be further expressed as:

[0132] φ Z (n,k)=α S (k)φ S (n,k)+α R (k)φ R (n,k)+α V (k)φ V (n,k);

[0133] in,

[0134]

[0135] The optimal Wiener gain is the ratio of the expected signal variance to the total noise variance. The present invention regards both reverberation and noise as unwanted noise. According to the above formula, the optimal Wiener gain G(n,k) can be expressed as:

[0136]

[0137] It should be noted that the distortion-free constraint is usually required to keep the desired source signal unattenuated, i.e., h H (k)d(k)=1.

[0138] Therefore, assuming that α S (k)≈1 is reasonable, so the above Wiener gain can be simplified to:

[0139]

[0140] in,

[0141]

[0142] GDR(n,k) and SNR(n,k) are the values ​​of CDR and SNR at the kth frequency point and the nth frame respectively.

[0143] The signal-to-noise ratio estimation value obtained in step S21 and the coherent scatter ratio estimation value obtained in step S22 are substituted into the Wiener gain function to calculate the Wiener gain.

[0144] In addition, in order to achieve a flexible trade-off between noise reduction and dereverberation, the present invention also adds two hyperparameters for balancing the attenuation degree of noise and reverberation in the Wiener gain. After adding the hyperparameters, the Wiener gain is further generalized as follows:

[0145]

[0146] Among them, β1≥0, β2≥0 are two hyperparameters used to control the trade-off between noise reduction and dereverberation attenuation. For example, if β2=1, a larger β1 can help more reverberation attenuation; if β1=1, a larger β2 corresponds to more noise reduction.

[0147] As we all know, there is often a trade-off between noise reduction and signal distortion. By applying the Wiener gain to the signal in the short-time Fourier transform (STFT) domain, more noise reduction usually means greater signal distortion. Of all types of signal distortion, musical noise is the most uncomfortable one. In order to prevent the filtered signal from being affected by musical noise, the Wiener gain actually applied to the observed signal is usually modified by the following formula:

[0148] G(n,k)←max{G(n,k),G min};

[0149] Among them, G min ∈(0,1) is a small positive number, such as G min = 0.01. Since the musical noise is caused by isolated peaks in the STFT domain, increasing G min can help reduce the difference between the value of isolated peaks and the background level, thereby reducing musical noise. However, using G min It also means less noise reduction, which is the price to pay for eliminating music noise.

[0150] The root cause of music noise is that the Wiener gain G(n,k) in some noise-dominated observation signals is large, or at least not small enough. According to the above Wiener gain expression, a larger G(n,k) requires both a larger SNR(n,k) and a larger CDR(n,k). If the values ​​of SNR and CDR are very large in certain frames and frequency points, then it is very likely that the observation signal is dominated by the source signal. In other words, if the observation signal is mainly dominated by noise and reverberation, it is unlikely to have a larger SNR(n,k) and a larger CDR(n,k), i.e., a larger G(n,k). Therefore, it can be expected that the method proposed in the present invention can also help reduce music noise. In practice, a smaller G can be set. min , that is, you can expect more noise reduction and dereverberation effects.

[0151] S3, processing the output signal through a post-filter to output a final source signal.

[0152] Specifically, the above post-filter G(n,k)∈[0,1] is applied to the output signal Z(n,k) of the beam, and the estimate of the final source signal can be expressed as:

[0153]

[0154] in, is the final source signal, G(n,k) is the Wiener gain, and Z(n,k) is the output signal of the beam.

[0155] In order to comprehensively evaluate the effect of the proposed speech enhancement technology, through a series of simulation experiments, the example of the present invention adopts three key indicators: signal-to-noise ratio gain (SNR Gain), log spectral distance (LSD) and direct mixing ratio (DRR). These indicators measure the performance of the system in speech noise reduction, distortion control and de-reverberation capabilities from different angles. Among them, the signal-to-noise ratio gain is defined as the ratio of the output signal-to-noise ratio to the input signal-to-noise ratio. The signal-to-noise ratio gain reflects the ability of the system to effectively improve the quality of speech signals and reduce noise interference. A high signal-to-noise ratio gain means that the system can significantly improve the clarity and intelligibility of speech signals. The log spectral distance refers to the difference between the expected signal and the output signal of the array in the log spectrum. The log spectral distance reflects the degree of distortion of the array output signal. A smaller log spectral distance indicates that the output signal is closer to the expected signal in the spectrum, that is, the distortion is lower. The direct-to-mix ratio is an important parameter to measure the relative strength between direct sound and reverberation sound. The direct-to-mix ratio reflects the system's ability to remove reverberation. A higher direct-to-mix ratio means that the system can effectively separate the direct sound and reduce the interference of reverberation sound, thereby improving the clarity and intelligibility of the observed signal.

[0156] The proposed method is compared with five other methods, including the traditional SNR-based Wiener method, CDR-based method, two-step noise reduction (TSNR), harmonic regeneration noise reduction (HRNR), and adaptive weighted predictive estimation (AWPE). In order to compare the performance of different methods fairly, all gains except AWPE are applied to the output of the super-directional (SD) beamformer. The resulting methods are called SD-SNR, SD-CDR, SD-TSNR, SD-HRNR, and AWPE.

[0157] (1) Impact of hyperparameters on performance

[0158] According to the designed Wiener gain, the present invention has two hyperparameters β1 and β2 that need to be determined. Figure 3a The trade-off between SNR Gain and DRR of the proposed method is shown as a function of β1 and β2, where the input SNR is 5dB and the reverberation time is 500ms. β1 and β2 vary from 1 to 10, increasing by 1 each time, and each curve corresponds to the same β2 value.

[0159] from Figure 3a As can be seen from the figure, as β1 increases, DRR also increases, and at the same time, as β2 increases, SNR Gain also increases. However, when β2 is greater than a certain value (for example, β2 ≥ 5), further increasing β2 does not greatly improve SNR Gain. In contrast, increasing β1 to improve DRR seems to be a more effective method. Figure 3a It can be seen from the figure that, compared with the traditional method, the method of the present invention is superior to the traditional method in terms of DRR and SNR Gain by taking appropriate β1 and β2.

[0160] Figure 3bThe trade-off between SNR Gain and LSD is shown as a function of β1 and β2, where β1 and β2 also vary from 1 to 10, increasing by 1 each time. First, as β2 increases, LSD decreases rapidly. Considering that SNR Gain increases with increasing β2, LSD will also decrease with increasing SNR Gain. However, by further increasing β2, when SNR gain increases, LSD will also increase, that is, more SNR improvement means greater signal distortion. Therefore, in practical situations, it may be a good strategy to increase β2 appropriately. In the case of β2>3, we can see that when β1 increases, SNR Gain decreases slightly, but DRR increases significantly, that is, we get more reverberation attenuation; however, as β1 increases, LSD also increases, that is, we have more signal distortion. Therefore, in practical applications, the value of β1 needs to be chosen based on the specific situation.

[0161] (2) Performance as a function of input signal-to-noise ratio

[0162] The performance of the proposed method under different input signal-to-noise ratios is simulated and evaluated, where the reverberation time is about 500ms and the additive noise signal consists of interference noise, isotropic noise and white noise. The SNR Gain, DRR and LSD of different methods are respectively Figure 4a , 4b and 4c. As can be seen from the figure, the proposed method outperforms the traditional method in terms of SNR Gain and LSD. Since the SNR Gain of SD beamforming is almost a constant under the input SNR, this is because its SNR Gain depends on the pseudo-coherence matrix of the noise environment, rather than the covariance matrix related to the noise variance.

[0163] However, due to the limitations of the array aperture and the number of sensors, the performance of the SD beamformer is very limited compared to other methods. Interestingly, the SD-CDR method can achieve the highest DRR, which means the best reverberation attenuation. In addition, as the input SNR increases, the SNR Gain decreases, that is, if the input SNR is relatively high, it will be more difficult to further improve the SNR. The LSD also decreases with the increase of the input SNR, which means that the observed signal with a higher input SNR has less distortion. Interestingly, for both the SD-CDR method and the method proposed in this invention, the change in DRR is very small when the input SNR changes. The reason behind this is that the Wiener gain of both methods is a function of CDR, which reflects the reverberation level, while the estimation of CDR is insensitive to the background noise level.

[0164] (3) Performance as a function of reverberation time

[0165] The proposed Wiener gain is a function of SNR and CDR. Considering that CDR reflects the reverberation level, it is interesting to study the performance of the proposed method under different reverberation conditions. The results are shown in Figure 5a , 5b and 5c, where the input SNR is set to 5dB. As we can see, the reverberation level has little effect on LSD, and only slightly reduces it under high reverberation conditions, because the signal distortion mainly comes from noise suppression. It can be seen that the DRR of all methods decreases rapidly with the increase of reverberation level. However, the DRR of the proposed method is larger than that of the traditional method, which is demonstrated by the local magnified views, where T60 is 200ms and 500ms respectively. In summary, under different reverberation conditions, the proposed method outperforms the traditional method in terms of SNR Gain, LSD and DRR.

[0166] In addition to CDR-based methods, WPE and AWPE methods are also widely used for dereverberation. In order to further compare the dereverberation performance of AWPE and other methods, since the DRR of AWPE cannot be obtained, the SRMR (speech-to-reverberation modulation energy ratio) of all methods is calculated. The results are shown in Figure 6 In the figure, the reverberation time T60 is set to 500ms. It can be observed that when the input SNR is high (e.g., input SNR=20dB), the SRMR of the AWPE method is slightly lower than that of the SD-CDR and the method proposed in the present invention. However, as the input SNR decreases, the SRMR of the AWPE method decreases rapidly, indicating that the AWPE method is very sensitive to additive noise. In contrast, it is obvious that the performance of the method proposed in the present invention is better than that of the AWPE method in a noisy and reverberant environment.

[0167] (4) Performance as the minimum filter gain G min Function

[0168] The minimum value of Wiener gain G min is an important hyperparameter in the filtering implementation. min Usually means a larger SNR gain. However, if G min If the value is less than a threshold, music noise will appear. In order to evaluate the degree of music noise, the kurtosis ratio (KR) is introduced. The kurtosis ratio can be defined as:

[0169]

[0170] Among them, Kurt out , kurtin They represent the fourth-order statistics kurtosis of the array output signal and the observed signal respectively. The kurtosis of a random process X can be defined as Here μ n Is Given the nth moment, P X (x) is the probability density function.

[0171] A larger kurtosis ratio means a higher risk of musical noise. The kurtosis ratio of the proposed method is used as G min Function, where the input signal-to-noise ratio is 5dB, T60 is about 500ms, and the result is as follows Figure 7 shown.

[0172] like Figure 7 As shown, with G min As G decreases, the kurtosis ratio will increase. min ≥0.1, it is found that the kurtosis ratio does not change much, so take G min = 0.1 seems to be a good choice. It can also be seen that the kurtosis ratio of the proposed method is smaller than that of the SD-SNR, SD-TSNR and SD-HRNR methods. Since SD is a fixed beamformer, musical noise cannot be introduced through this beamformer. It can also be seen that its kurtosis ratio remains very low. Interestingly, when G min When the change occurs, the kurtosis ratio of the SD-CDR method is also very small. However, considering that the SNR gain of the SD and SD-CDR methods is poor, the method proposed in the present invention is still better than the traditional method.

[0173] By using the above indicators in combination, the example of the present invention can comprehensively and objectively evaluate the performance of the proposed speech enhancement technology in different aspects. This multi-dimensional evaluation method not only helps to verify the effectiveness of the technology, but also provides a strong basis for further optimization and improvement.

[0174] The advantages of the present invention are as follows: 1. Aiming at the coexistence problem of noise and reverberation, the present invention proposes a novel beamformer design method, which adopts the classic beamforming + post-filtering framework, comprehensively considers CDR and SNR to design Wiener gain, and greatly improves the performance of speech enhancement technology in speech noise reduction, distortion control, and dereverberation capability. 2. The present invention also adopts two hyperparameters to control the attenuation degree of noise and reverberation, so as to flexibly adjust the balance between noise reduction and reverberation suppression, so as to improve the signal-to-noise ratio and reduce noise interference.

[0175] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A design method for Wiener gain, characterized in that: The method comprises: S1, in an environment where noise and reverberation coexist, processing an observation signal collected by a microphone array through a beamformer to output a beam output signal, wherein the observation signal includes a source signal and a noise source, and the noise source includes at least an additive noise signal and a reverberation signal; S2, performing Wiener gain estimation, and using the estimated Wiener gain as a post filter, the Wiener gain estimation process includes: S21, performing signal-to-noise ratio estimation according to the output signal in step S1 to obtain a signal-to-noise ratio estimation value; S22, performing coherent scatter ratio estimation according to the signal-to-noise ratio estimation value and the observation signal collected in step S1 to obtain a coherent scatter ratio estimation value; S23, performing gain estimation according to the estimated value of the signal-to-noise ratio and the estimated value of the coherent scattering ratio to obtain the Wiener gain; S3, processing the output signal through a post-filter to output a final source signal.

2. A method for designing Wiener gain as claimed in claim 1, characterized in that: In step S1, the observation signal is expressed as: y(n,k)=[Y1(n,k)Y2(n,k)···Y M (n,k)] T =d(k)S(n,k)+r(n,k)+v(n,k); Among them, y(n,k) is the observed signal vector, Y m (n,k) is the observation signal collected by the mth microphone, m = 1, 2, ···M, M is the number of elements in the microphone array, [·] T represents the transpose operation of a vector or matrix, d(k) represents the steering vector of the array, S(n,k) represents the source signal, r(n,k) represents the reverberation signal vector in the observed signal, v(n,k) represents the additive noise signal vector in the observed signal, n is the frame index, k=2πf represents the angular frequency, and f represents the frequency; The output signal is expressed as: z(n,k)=h H (k)y(n,k); Wherein, z(n,k) is the output signal vector, h(k) represents the weight vector of the beamformer, [·] H represents transposed conjugate.

3. A method for designing Wiener gain as claimed in claim 1, characterized in that: In S21, a direct decision method is used to estimate the signal-to-noise ratio.

4. A method for designing Wiener gain as claimed in claim 3, characterized in that: The calculation formula of the signal-to-noise ratio estimation value is: SNR(n,k)=αSNR(n-1,k)+(1-α)·max(ζ(n,k)-1,0); Among them, α∈(0,1) is a hyperparameter, n is a frame index, k=2πf represents the angular frequency, f represents the frequency, ζ represents the posterior signal-to-noise ratio, and ζ(n),k) is expressed as: Among them, z(n,k) is the output signal vector, φ V (n,k) is the variance of the additive noise signal.

5. A method for designing Wiener gain as claimed in claim 1, characterized in that: In S22, the calculation process of the coherent scatter ratio estimation value is: Calculate the estimated value of the coherent scattering ratio of the microphone pair according to the pseudo-coherence matrix and the estimated value of the signal-to-noise ratio of the observation signal, the additive noise signal and the reverberation signal collected by the microphone pair of any two channels in the microphone array; The coherent scatter ratio estimates of all microphone pairs in the microphone array are then averaged.

6. A method for designing Wiener gain as claimed in claim 5, characterized in that: The calculation formula of the coherent scatter ratio estimation value is: Where M is the number of elements in the microphone array, CDR i,j (k) is the estimated value of the coherent scatter ratio of the microphone pair, i = 1, 2 ... M-1, j = 0, 1 ... i-1, k = 2πf represents the angular frequency, and f represents the frequency; The coherent scatter ratio estimate for each microphone pair is given by: in, [Γ Y (k)] i,j is the pseudo-coherence matrix of the observation signal collected by the microphone pair, [Γ V (k)] i,j is the pseudo-coherence matrix of the additive noise signal collected by the microphone pair, [Γ R (k)] i,j is the pseudo-coherence matrix of the reverberation signal collected by the microphone pair, represents the real part operation, [·] * represents the conjugate transpose, and SNR(k) is the signal-to-noise ratio estimate.

7. A method for designing Wiener gain as claimed in claim 1, characterized in that: In S23, the calculation process of the Wiener gain includes: Constructing a Wiener gain function based on a coherent spread ratio and a signal-to-noise ratio from the output signal of the beam in step S1; The signal-to-noise ratio estimation value obtained in step S21 and the coherent scatter ratio estimation value obtained in step S22 are substituted into the Wiener gain function to calculate the Wiener gain.

8. A method for designing Wiener gain as claimed in claim 7, characterized in that: In S23, the Wiener gain is expressed as: in, h(k) represents the weight vector of the beamformer, [·] H represents the transposed conjugate, Γ R (k),Γ V (k) represents the pseudo-coherence matrix of the reverberation signal vector and the additive noise signal vector, respectively, CDR(n,k) represents the estimated value of the coherent scattering ratio, and SNR(n,k) represents the estimated value of the signal-to-noise ratio.

9. A method for designing Wiener gain as claimed in claim 8, characterized in that: The S2 further includes: S24, adding two hyperparameters for balancing the attenuation degree of noise and reverberation to the Wiener gain, and after adding the hyperparameters, the Wiener gain is expressed as: Among them, β1≥0, β2≥0 are two hyperparameters.

10. A method for designing Wiener gain according to claim 1 or 6, characterized in that: In S3, the final source signal is expressed as: in, is the final source signal, G(n,k) is the Wiener gain, Z(n,k) is the output signal of the beam, n is the frame index, k=2πf represents the angular frequency, and f represents the frequency.

Images