A voice signal processing method and device and electronic equipment

Abstract

Embodiments of the present application disclose a voice signal processing method and device and electronic equipment. The device comprises a dereverberation module and a blind source separation module; wherein: the dereverberation module is configured to utilize a prediction coefficient matrix G(n) of the n th moment to perform dereverberation processing on a delay signal group Y(n-D) corresponding to an input signal y(n) of the n th moment, to obtain a dereverberation actual signal x(n) of the n th moment; the blind source separation module is configured to utilize a separation matrix W(n) of the n th moment to process the dereverberation actual signal x(n) of the n th moment, to obtain a separation actual signal z(n) of the n th moment; wherein the device further comprises a calculation module configured to calculate the prediction coefficient matrix G(n) of the n th moment by the following method, comprising: obtaining a separation matrix W(n-1) of the (n-1) th moment; obtaining the prediction coefficient matrix G(n) of the n th moment according to the separation matrix W(n-1) of the (n-1) th moment and a prediction coefficient matrix G(n-1) of the (n-1) th moment.

Description

Technical Field

[0001] This application relates to the field of audio processing, and more particularly to a speech signal processing method, apparatus, and electronic device. Background Technology

[0002] Blind source separation (BSS) is a technique for separating multiple sound sources from an observed signal. "Blind" means that the source signals and transmission channels are unknown. Multi-channel blind source separation is categorized into three cases: overdetermined, positive definite, and underdetermined, based on whether the number of microphones is greater than, equal to, or less than the number of sound sources. Most blind source separation algorithms focus on the case where the number of microphones is greater than or equal to the number of sound sources. Since the overdetermined case can be transformed into a positive definite problem using Principal Component Analysis (PCA).

[0003] For blind source separation techniques in the positive definite case, commonly used methods include Independent Vector Analysis (ICA), Independent Vector Analysis (IVA), Multichannel Non-Negative Matrix Factorization (MNMF), and Independent Low-Rank Matrix Analysis (ILRMA). IVA, compared to ICA, overcomes the frequency ambiguity problem and has therefore received widespread attention. Auxiliary Function Based IVA (AuxIVA) has become a mainstream approach for solving BSS problems in recent years, and one such method provides a real-time IVA approach that updates the separation matrix frame by frame.

[0004] The signal captured by the microphone includes not only the target sound source but also signals from other interfering sound sources and room reverberation, which significantly degrade the performance of the audio system. Blind source separation technology can separate different sound sources, thereby eliminating the adverse effects of interfering sound sources on the target sound source. However, the performance of blind source separation technology is greatly affected by reverberation; as the room reverberation time increases, the performance of blind source separation rapidly declines. Therefore, the input signal should be pre-processed to remove reverberation.

[0005] The most common dereverberation techniques fall into two categories: noise suppression and linear prediction. Noise suppression methods first estimate the reverberation energy, treating it as noise, and then use noise suppression techniques such as spectral subtraction and Wiener filtering to remove the reverberation component. Linear prediction methods use data from several past frames to predict the reverberation component in the current speech frame, and then subtract it. Among these, the Weighted Prediction Error (WPE) method and its adaptive variant AWPE are widely used due to their excellent dereverberation performance.

[0006] In a reverberant environment, the reverberation generated by the reflection and refraction of sound waves can interfere with blind source separation and greatly reduce the separation performance. A technical solution combining dereverberation and blind source separation has been proposed. However, when performing dereverberation processing, the input signal is expanded by Kronecker product with the identity matrix, resulting in poor dereverberation effect. Summary of the Invention

[0007] To address any of the aforementioned technical problems, embodiments of this application provide a speech signal processing method, apparatus, and electronic device.

[0008] To achieve the objectives of the embodiments of this application, this application provides a speech signal processing apparatus, the apparatus including a de-reverberation module and a blind source separation module; wherein:

[0009] The dereverberation module is used to perform dereverberation processing on the delayed signal group Y(nD) corresponding to the input signal y(n) at time n using the prediction coefficient matrix G(n) at time n, so as to obtain the dereverberated actual signal x(n) at time n.

[0010] The blind source separation module is used to process the de-reverberation actual signal x(n) at time n using the separation matrix W(n) at time n to obtain the separated actual signal z(n) at time n.

[0011] The device further includes:

[0012] The calculation module is used to calculate the prediction coefficient matrix G(n) at time n in the following manner:

[0013] Obtain the separation matrix W(n-1) at time n-1;

[0014] Based on the separation matrix W(n-1) at time n-1 and the prediction coefficient matrix G(n-1) at time n-1, the prediction coefficient matrix G(n) at time n is obtained.

[0015] in:

[0016]

[0017] Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D and K are all positive integers.

[0018] A speech signal processing method, applied to the apparatus described above, the method comprising:

[0019] Obtain the separation matrix W(n-1) at time n-1;

[0020] Based on the separation matrix W(n-1) at time n-1 and the prediction coefficient matrix G(n-1) at time n-1, the prediction coefficient matrix G(n) at time n is obtained.

[0021] Using the prediction coefficient matrix G(n) at time n, the delayed signal group Y(nD) corresponding to the input signal y(n) at time n is de-reverberated to obtain the de-reverberated actual signal x(n) at time n.

[0022] The blind source separation module is used to process the de-reverberation actual signal x(n) at time n using the separation matrix W(n) at time n to obtain the separated actual signal z(n) at time n.

[0023] in:

[0024]

[0025] Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D and K are all positive integers.

[0026] An electronic device includes a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the methods described above.

[0027] One of the above technical solutions has the following advantages or beneficial effects:

[0028] By obtaining the separation matrix W(n-1) at time n-1, prior information of the speech signal is obtained, and the prediction coefficient matrix G(n) at time n is generated based on this prior information. The result is more accurate and can suppress the reverberation of the speech signal more effectively, further improving the performance of blind source separation in high reverberation scenarios.

[0029] Other features and advantages of the embodiments of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the embodiments of this application. The objects and other advantages of the embodiments of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0030] The accompanying drawings are used to provide a further understanding of the technical solutions of the embodiments of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions of the embodiments of this application.

[0031] Figure 1 A schematic diagram of the speech signal processing apparatus 100 provided in an embodiment of this application;

[0032] Figure 2 for Figure 1 A schematic diagram of the structure of the de-reverberation module 10 in the device 100 shown;

[0033] Figure 3 for Figure 1 Another schematic diagram of the device 100 shown;

[0034] Figure 4 for Figure 3 Another schematic diagram showing setting 100;

[0035] Figure 5 A flowchart of a speech signal processing method provided in an embodiment of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0037] Figure 1 This is a schematic diagram of a speech signal processing apparatus 100 provided in an embodiment of this application. Figure 1 As shown, the device 100 includes a de-reverberation module 10 and a blind source separation module 20; wherein, y(n) = (y1(n), y2(n)..., y M (n)) T The input signal represents multiple signals collected by the microphone, x(n) = (x1(n), x2(n), ..., x...). M (n)) T This represents the deveraging result output by deveraging module 10, z(n) = (z1(n), z2(n), ..., z M (n)) T This indicates the blind source separation result output by the blind source separation module.

[0038] The de-reverberation module 10 is used to perform de-reverberation processing on the delayed signal group Y(nD) corresponding to the input signal y(n) at time n using the prediction coefficient matrix G(n) at time n, so as to obtain the de-reverberation actual signal x(n) at time n.

[0039] in:

[0040]

[0041]

[0042] The initial value of the prediction coefficient matrix G(n) can be set to all zeros;

[0043] Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D and K are all positive integers.

[0044] Specifically, the dereverberation process can be represented as:

[0045] x(n)=y(n)-G H (n)Y(nD)

[0046] De-reverberation processing involves applying the above formula frame by frame, starting from time n = D + K, to subtract the estimated reverberation component G(n) from the output signal y(n). H Y(nD) yields the actual dereverberated signal x(n).

[0047] Figure 2 for Figure 1 A schematic diagram of the de-reverberation module 10 in the device 100 shown. Figure 2 As shown, the dereverberation module 10 includes K branches, wherein the k-th branch uses the k-th set of reverberation prediction coefficients to perform dereverberation processing on the input signal at time (nD-k+1), where k = 1, 2, 3, ..., K.

[0048] Since the k-th branch uses the k-th set of reverberation prediction coefficients to dereverberate the input signal at time (nD-k+1), the goal of parallel signal processing by the K branches is achieved, reducing the computational complexity of dereverberation.

[0049] The blind source separation module 20 is connected to the dereverberation module 10 and is used to process the dereverberation actual signal x(n) at time n using the separation matrix W(n) at time n to obtain the separation actual signal z(n) at time n.

[0050] The blind source separation module 20 can use various algorithms such as ICA, AuxIVA, MNMF, and ILRMA to determine the separation matrix W(n) at time n, which will not be elaborated here.

[0051] After obtaining the separation matrix W(n) at time n, the blind source separation module 20 can obtain the separated actual signal z(n) at time n through the following calculation expression, including:

[0052] z(n) = W(n)x(n).

[0053] The apparatus provided in this application combines dereverberation with blind source separation. By effectively suppressing the reverberation of the input signal before performing blind source separation, the performance of blind source separation in high reverberation scenarios can be effectively improved.

[0054] Figure 3 for Figure 1 Another schematic diagram of the device 100 shown. (As shown) Figure 3 As shown, the device 100 further includes:

[0055] The calculation module 30, connected to the dereverberation module 10 and the blind source separation module 20, is used to calculate the prediction coefficient matrix G(n) at time n in the following manner:

[0056] Obtain the separation matrix W(n-1) at time n-1;

[0057] Based on the separation matrix W(n-1) at time n-1 and the prediction coefficient matrix G(n-1) at time n-1, the prediction coefficient matrix G(n) at time n is obtained.

[0058] Specifically, the calculation module 30 obtains the separation matrix W(n-1) at time n-1 from the blind source separation module 20 to obtain prior information of the speech signal, and obtains the prediction coefficient matrix G(n) at time n based on the separation matrix W(n-1) at time n-1 and the prediction coefficient matrix G(n-1) at time n-1. This makes the prediction coefficient matrix G(n) at time n generated based on prior information. Compared with generating the reverberation prediction coefficient matrix G(n) at time n without using prior information, the determination result of the prediction coefficient matrix G(n) at time n provided in this application embodiment is more accurate.

[0059] The apparatus provided in this application obtains prior information of the speech signal by acquiring the separation matrix W(n-1) at time n-1, and generates the prediction coefficient matrix G(n) at time n based on the prior information. The determination result is more accurate, and the reverberation of the speech signal can be suppressed more effectively, further improving the performance of blind source separation in high reverberation scenarios.

[0060] In an exemplary embodiment, the dereverberation module 10 is further configured to process the input signal y(n) at time n using the prediction coefficient matrix G(n-1) at time n-1 to obtain the dereverberation prediction signal x at time n. p (n);

[0061] Specifically, since the prediction coefficient matrix G(n) at time n is unknown, the prediction coefficient matrix from the previous time step can be used as an approximation to complete the dereverberation prediction signal x at time n. p Obtaining (n).

[0062] The blind source separation module 20 is further configured to use the separation matrix W(n-1) at time n-1 to perform a de-reverberation prediction on the denoised prediction signal x at time n. p (n) is processed to obtain the separation prediction signal z at time n. p (n);

[0063] Specifically, since the separation matrix W(n) at time n is unknown, the separation matrix from the previous time step can be used as an approximation to complete the separation prediction of the signal z at time n. p Obtaining (n).

[0064] The calculation module 30 is used to perform a separation prediction signal z at time n. p (n), calculate the average power λ at time n. n And based on the average power λ at time n n The prediction coefficient matrix G(n-1) at time n-1 is processed to obtain the prediction coefficient matrix G(n) at time n.

[0065] Specifically, due to the separation prediction signal z at time n p z(n) is the predicted source signal's speech data, which has a strong similarity to the actual separated signal z(n) at time n. Therefore, the predicted separated signal z(n) at time n can be used. p The prior information of the speech signal carried in (n) is used to generate the prediction coefficient matrix G(n) at time n.

[0066] Furthermore, the separation prediction signal z at time n is used. p (n) is used to calculate the average power λ at time n. n This allows full utilization of the separation prediction signal z at time n. p (n) Prior information of the speech signal carried.

[0067] Compared with the dereverberation prediction signal x at time n p (n) is used to calculate the average power λ at time n.n In comparison, the device provided in this application calculates the average power λ at time n. n The parameters used are different.

[0068] Specifically, due to the separation prediction signal z at time n p (n) represents the dereverberation prediction signal x at time n. p (n) is the signal obtained after processing. This is compared with the dereverberation prediction signal x at time n. p Compared to (n), the separation prediction signal z at time n. p (n) is the source signal at time n obtained based on the prediction operation, while the de-reverberation prediction signal x at time n is obtained at time n. p (n) represents multiple signals obtained from the prediction operation, including the source signal. This is due to the separation prediction signal z at time n. p (n) is unmixed with other signals; therefore, due to the separation prediction signal z p (n) carries more accurate speech information from the source signal, resulting in a higher average power λ at time n. n The determination results are more accurate, providing support for more accurate calculation of the prediction coefficient matrix G(n) at time n.

[0069] In an exemplary embodiment, the calculation module 30 calculates the prediction coefficient matrix G(n) at time n in the following manner:

[0070] Based on the average power λ at time n n Determine the adjustment coefficient k(n) at time n;

[0071] Calculate the adjustment coefficient k(n) at time n and the dereverberation prediction signal x at time n. p The transpose conjugate matrix of (n) The product of these two elements yields the adjustment matrix at time n.

[0072] The prediction coefficient matrix G(n-1) at time n-1 is updated using the adjustment matrix at time n-1 to obtain the prediction coefficient matrix G(n) at time n.

[0073] Specifically, the calculation module 30 calculates the adjustment coefficient k(n) at time n in the following manner:

[0074]

[0075] in,

[0076] Wherein, γ is the forgetting factor, and the value of γ is between 0.9 and 1.

[0077] Specifically, the calculation module 30 calculates the prediction coefficient matrix G(n) at time n in the following manner:

[0078]

[0079] The above method can fully utilize the calculated dereverberation prediction signal x at time n. p (n) and the separation prediction signal z at time n p (n), obtain the adjustment information corresponding to the prediction coefficient matrix G(n-1) at time n-1, and be able to determine the prediction coefficient matrix G(n) at time n more accurately.

[0080] In one exemplary embodiment, the calculation module 30 obtains the average power λ at time n in the following manner. n ,include:

[0081] Obtain the sum of the diagonal elements from the covariance matrix of the separated prediction signals of the M channels at time n;

[0082] Based on the sum of the diagonal elements and the M channels, the average power λ at time n is obtained. n .

[0083] Specifically, the covariance matrix of the separated prediction signals of the M channels at time n. The expression is as follows:

[0084]

[0085] Among them, z p,m (n) represents the separation prediction signal of the m-th channel at time n, where m = 1, 2, 3, ..., M.

[0086] Specifically, the calculation module 30 obtains the average power λ at time n through the following calculation expression. n .

[0087]

[0088] in, Represents the covariance matrix The sum of the elements along the diagonal.

[0089] Figure 4 for Figure 3 Another schematic diagram of the device 100 shown. (As shown) Figure 4 As shown, the device further includes a first conversion module 40 and a second conversion module 50; wherein:

[0090] The first conversion module 40 is connected to the de-reverberation device 10 and is used to convert the received audio signal from a time domain signal to a frequency domain signal.

[0091] The second conversion module 50 is connected to the blind source separation module 20 and is used to convert the received blind source separation result from the frequency domain signal to the time domain signal.

[0092] Since the acquired speech signal is a short-term stable time-domain signal, the first conversion module 40 converts the time-domain signal to a frequency-domain signal and then outputs it to the dereverberation device 10. After the dereverberation device 10 and the blind source separation module 20 complete the processing of the frequency-domain signal, the second conversion module 50 converts the time-domain signal to a frequency-domain signal for each frame of speech signal.

[0093] The first conversion module 40 converts the audio signal from a time-domain signal to a frequency-domain signal in the following manner:

[0094] The acquired audio signal is divided into frames, where the duration of each frame can be from 10ms to 30m, and a certain overlap ratio is set.

[0095] Choose a time-domain window function, and apply a window to the time-domain signal frame by frame by moving the time-domain window function. The time-domain window function can be a Hanning window.

[0096] Perform a Fast Fourier Transform (FFT) on each frame of signal to convert the time-domain signal to the frequency domain.

[0097] The second conversion module 50 converts the blind source separation result from the frequency domain signal to the time domain signal in the following manner:

[0098] The blind source separation results of each frame are subjected to Inverse Fast Fourier Transform (IFFT) to convert the frequency domain signal into the time domain signal.

[0099] Extract the real part from the time-domain signal and perform windowing processing;

[0100] The target audio is obtained by adding the signals according to the set repetition rate.

[0101] Figure 5 A flowchart illustrating a speech signal processing method provided in an embodiment of this application. Figure 5 As shown, the method is applied to Figure 3 and Figure 4 The apparatus shown, the method includes:

[0102] Step A: Obtain the separation matrix W(n-1) at time n-1;

[0103] Step B: Based on the separation matrix W(n-1) at time n-1 and the prediction coefficient matrix G(n-1) at time n-1, obtain the prediction coefficient matrix G(n) at time n.

[0104] Step C: Using the prediction coefficient matrix G(n) at time n, perform dereverberation processing on the delayed signal group Y(nD) corresponding to the input signal y(n) at time n to obtain the dereverberated actual signal x(n) at time n.

[0105] Step D: Using the separation matrix W(n) at time n, process the dereverberated actual signal x(n) at time n to obtain the separated actual signal z(n) at time n.

[0106] in:

[0107]

[0108] Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D and K are all positive integers.

[0109] In one exemplary embodiment, step A includes:

[0110] Step A1: Using the prediction coefficient matrix G(n-1) at time n-1, process the input signal y(n) at time n to obtain the dereverberation prediction signal x at time n. p (n);

[0111] Specifically, since the prediction coefficient matrix G(n) at time n is unknown, the prediction coefficient matrix from the previous time step can be used as an approximation to complete the dereverberation prediction signal x at time n. p Obtaining (n).

[0112] Step A2: Using the separation matrix W(n-1) at time n-1, predict the dereverberation signal x at time n. p (n) is processed to obtain the separation prediction signal z at time n. p (n);

[0113] Specifically, since the separation matrix W(n) at time n is unknown, the separation matrix from the previous time step can be used as an approximation to complete the deseparation prediction of the signal z at time n. p Obtaining (n).

[0114] Step A3: Based on the separation prediction signal z at time n p (n), calculate the average power λ at time n. n ;

[0115] Due to the separation prediction signal z at time n p z(n) is the predicted speech data after removing interference. It has a strong similarity to the actual separated signal z(n) at time n. Therefore, the predicted separated signal z(n) at time n can be used. p The prior information of the speech signal carried in (n) is used to calculate the prediction coefficient matrix G(n) at time n.

[0116] Compared with the dereverberation prediction signal x at time n p (n) is used to calculate the average power λ at time n. n In comparison, the device provided in this application calculates the average power λ at time n. n The parameters used are different.

[0117] Specifically, due to the separation prediction signal z at time n p (n) represents the dereverberation prediction signal x at time n. p (n) is the signal obtained after processing. This is compared with the dereverberation prediction signal x at time n. p Compared to (n), the separation prediction signal z at time n. p (n) is the source signal at time n obtained based on the prediction operation, while the de-reverberation prediction signal x at time n is obtained at time n. p (n) represents multiple signals obtained from the prediction operation, including the source signal. This is due to the separation prediction signal z at time n. p (n) is unmixed with other signals; therefore, due to the separation prediction signal z p (n) carries more accurate speech information from the source signal, resulting in a higher average power λ at time n. n The determination results are more accurate, providing support for more accurate calculation of the prediction coefficient matrix G(n) at time n.

[0118] Step A4: Based on the average power λ at time n n The prediction coefficient matrix G(n-1) at time n-1 is processed to obtain the prediction coefficient matrix G(n) at time n.

[0119] Specifically, the adjustment coefficient k(n) at time n is calculated as follows:

[0120]

[0121] in,

[0122] Wherein, γ is the forgetting factor, and the value of γ is between 0.9 and 1.

[0123] In one exemplary embodiment, step A4 includes:

[0124] Step A41: Based on the average power λ at time n n Determine the adjustment coefficient k(n) at time n;

[0125] Step A42: Calculate the adjustment coefficient k(n) at time n and the dereverberation prediction signal x at time n. p The transpose conjugate matrix of (n) The product of these two elements yields the adjustment matrix at time n.

[0126] Step A43: Using the adjustment matrix at time n, update the prediction coefficient matrix G(n-1) at time n-1 to obtain the prediction coefficient matrix G(n) at time n.

[0127] Specifically, the calculation module calculates the prediction coefficient matrix G(n) at time n in the following manner:

[0128]

[0129] The above method can fully utilize the calculated dereverberation prediction signal x at time n. p (n) and the separation prediction signal z at time n p (n), obtain the adjustment information corresponding to the prediction coefficient matrix G(n-1) at time n-1, and be able to determine the prediction coefficient matrix G(n) at time n more accurately.

[0130] In one exemplary embodiment, step A3 includes:

[0131] Step A31: Obtain the sum of the diagonal elements from the covariance matrix of the separated prediction signals of the M channels at time n;

[0132] Specifically, the covariance matrix of the separated prediction signals of the M channels at time n. The expression is as follows:

[0133]

[0134] Among them, z p,m (n) represents the separation prediction signal of the m-th channel at time n, where m = 1, 2, 3, ..., M.

[0135] Step A32: Based on the sum of the diagonal elements and the M channels, obtain the average power λ at time n. n .

[0136] Specifically, the average power λ at time n is obtained through the following calculation expression. n .

[0137]

[0138] This indicates that, Represents the covariance matrix The sum of the elements along the diagonal.

[0139] The method provided in this application obtains prior information of the speech signal by acquiring the separation matrix W(n-1) at time n-1, and generates the prediction coefficient matrix G(n) at time n based on the prior information. This results in more accurate determination and can more effectively suppress the reverberation of the speech signal, further improving the performance of blind source separation in high reverberation scenarios.

[0140] This application provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the methods described above.

[0141] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

Claims

1. A speech signal processing device, characterized in that, The device includes a reverberation removal module and a blind source separation module; wherein: The dereverberation module is used to utilize the nth... Prediction coefficient matrix For the input signal at time n Corresponding delay signal group Perform dereverberation processing to obtain the actual dereverberated signal at time n. ; The blind source separation module is used to utilize the nth... Separation matrix For the dereverberated actual signal at time n The signal is processed to obtain the actual separated signal at time n. ; The device further includes: The calculation module is used to calculate the first... Prediction coefficient matrix ,include: Get the (n-1)th Separation matrix ; According to the (n-1)th Separation matrix The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. ; in: Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D, and K are all positive integers; wherein, the de-reverberation module is further used to utilize the prediction coefficient matrix at time n-1. For the input signal at time n The signal is processed to obtain the dereverberation prediction signal at time n. ; The blind source separation module is also used to utilize the separation matrix at time n-1. For the dereverberation prediction signal at time n The signal is processed to obtain the separation prediction signal at time n. ; The calculation module is used to predict the separation signal at time n. Calculate the average power at time n. And based on the average power at time n The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. .

2. The apparatus according to claim 1, characterized in that, The calculation module calculates the first... ,include: Based on the average power at time n Determine the adjustment coefficient at time n. ; Calculate the adjustment factor at time n. With the dereverberation prediction signal at time n transpose conjugate matrix The product of these two elements yields the adjustment matrix at time n. Using the adjustment matrix at time n, the prediction coefficient matrix at time n-1 is... Update to obtain the nth... .

3. The apparatus according to claim 2, characterized in that, The calculation module calculates the adjustment coefficient at time n in the following manner. ,include: in, in, Forgetting factor, The value is between 0.9 and 1.

4. The apparatus according to any one of claims 1 to 3, characterized in that, The calculation module obtains the average power at time n in the following manner. ,include: Obtain the sum of the diagonal elements from the covariance matrix of the separated prediction signals of the M channels at time n; Based on the sum of the diagonal elements and the M channels, the average power at time n is obtained. .

5. A speech signal processing method, characterized in that, Applied to the apparatus of any one of claims 1 to 4, the method comprises: Get the (n-1)th Separation matrix ; According to the (n-1)th Separation matrix The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. ; Using the nth Prediction coefficient matrix For the input signal at time n Corresponding delay signal group Perform dereverberation processing to obtain the actual dereverberated signal at time n. ; Using the nth Separation matrix For the dereverberated actual signal at time n The signal is processed to obtain the actual separated signal at time n. ; in: Where M represents the total number of microphone channels, D represents the minimum prediction time interval, K represents the total number of signals at different delay times in the delayed signal group, and n, M, D and K are all positive integers; Wherein, according to the (n-1)th Separation matrix The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. ,include: Using the prediction coefficient matrix at time n-1 For the input signal at time n The signal is processed to obtain the dereverberation prediction signal at time n. ; Using the separation matrix at time n-1 For the dereverberation prediction signal at time n The signal is processed to obtain the separation prediction signal at time n. ; Based on the separation prediction signal at time n Calculate the average power at time n. ; Based on the average power at time n The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. .

6. The method according to claim 5, characterized in that, The average power at time n The prediction coefficient matrix at time n-1 The prediction coefficient matrix at time n is obtained. ,include: Based on the average power at time n Determine the adjustment coefficient at time n. ; Calculate the adjustment factor at time n. With the dereverberation prediction signal at time n transpose conjugate matrix The product of these two elements yields the adjustment matrix at time n. Using the adjustment matrix at time n, the prediction coefficient matrix at time n-1 is... Update to obtain the nth time step. .

7. The method according to claim 6, characterized in that, The adjustment factor at time n is calculated as follows: ,include: in, in, Forgetting factor, The value is between 0.9 and 1.

8. The method according to any one of claims 5 to 7, characterized in that, Based on the separation prediction signal at time n Calculate the average power at time n. ,include: Obtain the sum of the diagonal elements from the covariance matrix of the separated prediction signals of the M channels at time n; Based on the sum of the diagonal elements and the M channels, the average power at time n is obtained. .

9. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method according to any one of claims 5 to 8.

Images