A method, apparatus, and storage medium for suppressing late reverberation

Abstract

This disclosure relates to the field of speech signal processing technology, and discloses a method, apparatus, and storage medium for suppressing late reverberation. The method involves: performing noise reduction processing on the original speech signal using a neural network, determining the 2-norm weighting coefficients based on the noise reduction result signal, inputting the 2-norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients, determining the late reverberation estimate based on the filter coefficients, wherein the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain, determining the difference between the speech signal to be processed and the late reverberation estimate as the target speech signal, determining the accurate 2-norm weighting coefficients through the neural network to obtain the accurate late reverberation estimate, and subtracting the late reverberation estimate from the speech signal to be processed to obtain the target speech signal, thereby achieving the effect of suppressing late reverberation.

Description

Technical Field

[0001] This disclosure relates to the field of speech signal processing technology, and provides a method, apparatus and storage medium for suppressing late reverberation. Background Technology

[0002] As more and more voice devices appear in various scenarios, such as electronic products that work in hands-free mode, voice recognition systems in video conferencing, automatic translation systems in remote conferencing, and interactive television, the voice signals collected by voice devices will inevitably be contaminated by reverberation and noise. If the reverberation (usually referring to late reverberation in the industry) is severe, it will cause the voice device to malfunction, that is, the voice signal cannot be played normally.

[0003] Existing methods for suppressing late reverberation mainly involve first dividing the speech signal into frames, then converting it to the frequency domain, directly estimating the late reverberation using an exponential decay model in the frequency domain, and then combining this with the QR-RLS algorithm to suppress the multi-channel reverberation signal. However, the above method of directly using an exponential decay model is too crude and simple, and the accuracy of the estimated late reverberation is not high. Although spectral subtraction with coefficients is added in some scenarios, this coefficient (usually referring to the 2-norm weighting coefficient in the industry) can only be adjusted manually, and the calculation process is time-consuming and laborious.

[0004] In summary, there is currently no effective solution to suppress late reverberation in speech signals under noisy conditions. Summary of the Invention

[0005] This disclosure provides a method, apparatus, and storage medium for suppressing late reverberation, thereby improving the suppression effect of late reverberation in speech signals under noisy environments.

[0006] The specific technical solution provided in this disclosure is as follows:

[0007] In a first aspect, embodiments of this disclosure provide a method for suppressing late reverberation, comprising:

[0008] The original speech signal is denoised using a neural network, and the 2-norm weighting coefficients are determined based on the denoised signal.

[0009] The L2 norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal are input into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain the filter coefficients, and the late reverberation estimate is determined based on the filter coefficients. The speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain.

[0010] The difference between the speech signal to be processed and the late reverberation estimate is determined as the target speech signal.

[0011] Optionally, the 2-norm weighting coefficients of the reference signal are determined based on the denoised result signal, including:

[0012] The noise-reduced signal is converted from the time domain to the frequency domain to obtain the reference signal;

[0013] The power spectral density is determined based on the amplitude value of the reference signal;

[0014] The 2-norm weighting coefficients of the reference signal are determined based on the average power spectral density, the number of channels of the reference signal, and the morphological parameters, where the number of channels and the morphological parameters are attributes of the reference signal.

[0015] Optionally, the noise-reduced signal is transformed from the time domain to the frequency domain to obtain a reference signal, including:

[0016] The noise-reduced signal is segmented in the time domain according to a specified step size to obtain multiple segmented signal results.

[0017] The segmented result signals are multiplied by a preset window to obtain multiple window signals;

[0018] Perform a short-time Fourier transform on each window signal to obtain multiple sub-reference signals in the frequency domain;

[0019] The frequency domain signal corresponding to the result signal is determined based on each sub-reference signal, and the frequency domain signal corresponding to the result signal is used as the reference signal.

[0020] Optionally, the speech signal to be processed corresponding to the original speech signal is determined in the following way:

[0021] The original speech signal is segmented in the time domain according to a specified step size to obtain multiple original segmented result signals;

[0022] The original segmented result signals are multiplied by the preset window to obtain multiple original window signals;

[0023] The original window signals are subjected to short-time Fourier transform to obtain multiple sub-original reference signals in the frequency domain. The frequency domain signal corresponding to the original speech signal is determined based on each sub-original reference signal, and the frequency domain signal corresponding to the original speech signal is used as the speech signal to be processed.

[0024] Optionally, determining the late reverberation estimate based on the filter coefficients includes:

[0025] Transpose the filter coefficients to obtain the transposed filter coefficients;

[0026] Based on the delay time of the reference signal, a signal buffer block is determined, wherein the signal buffer block includes the early reverberation of the reference signal;

[0027] Based on the transposed filter coefficients and the signal buffer block, the late reverberation estimate is determined.

[0028] Secondly, embodiments of this disclosure also provide an apparatus for suppressing late reverberation, comprising:

[0029] The coefficient determination unit is used to perform noise reduction processing on the original speech signal through a neural network, and to determine the 2-norm weight coefficients based on the result signal after noise reduction.

[0030] The late reverberation determination unit is used to input the 2-norm weight coefficients and the speech signal to be processed corresponding to the original speech signal into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients, and determine the late reverberation estimate based on the filter coefficients. The speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain.

[0031] The speech determination unit is used to determine the difference between the speech signal to be processed and the late reverberation estimate as the target speech signal.

[0032] Optionally, the 2-norm weighting coefficients are determined based on the denoised signal, and the coefficient determination unit is used for:

[0033] The noise-reduced signal is converted from the time domain to the frequency domain to obtain the reference signal;

[0034] The power spectral density is determined based on the amplitude value of the reference signal;

[0035] The 2-norm weighting coefficients of the reference signal are determined based on the average power spectral density, the number of channels of the reference signal, and the morphological parameters, where the number of channels and the morphological parameters are attributes of the reference signal.

[0036] Optionally, the noise-reduced signal is transformed from the time domain to the frequency domain to obtain a reference signal, which is then used by the late reverberation determination unit:

[0037] The noise-reduced signal is segmented in the time domain according to a specified step size to obtain multiple segmented signal results.

[0038] The segmented result signals are multiplied by a preset window to obtain multiple window signals;

[0039] Perform a short-time Fourier transform on each window signal to obtain multiple sub-reference signals in the frequency domain;

[0040] The frequency domain signal corresponding to the result signal is determined based on each sub-reference signal, and the frequency domain signal corresponding to the result signal is used as the reference signal.

[0041] Optionally, the speech signal to be processed corresponding to the original speech signal is determined in the following way:

[0042] The original speech signal is segmented in the time domain according to a specified step size to obtain multiple original segmented result signals;

[0043] The original segmented result signals are multiplied by the preset window to obtain multiple original window signals;

[0044] The original window signals are subjected to short-time Fourier transform to obtain multiple sub-original reference signals in the frequency domain. The frequency domain signal corresponding to the original speech signal is determined based on each sub-original reference signal, and the frequency domain signal corresponding to the original speech signal is used as the speech signal to be processed.

[0045] Optionally, the late reverberation estimate is determined based on the filter coefficients, and the late reverberation determination unit is used for:

[0046] Transpose the filter coefficients to obtain the transposed filter coefficients;

[0047] Based on the delay time of the reference signal, a signal buffer block is determined, wherein the signal buffer block includes the early reverberation of the reference signal;

[0048] Based on the transposed filter coefficients and the signal buffer block, the late reverberation estimate is determined.

[0049] Thirdly, a smart terminal includes:

[0050] Memory, used to store executable instructions;

[0051] A processor for reading and executing executable instructions stored in memory to implement the method as described in any of the first aspects.

[0052] Fourthly, a computer-readable storage medium, when instructions in the storage medium are executed by a processor, enables the processor to perform the method described in any of the first aspects above.

[0053] The beneficial effects of this disclosure are as follows:

[0054] In summary, the embodiments of this disclosure provide a method, apparatus, and storage medium for suppressing late reverberation. The method includes: performing noise reduction processing on the original speech signal using a neural network, determining the 2-norm weighting coefficients based on the noise reduction result signal, inputting the 2-norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal into an RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients, and determining a late reverberation estimate based on the filter coefficients. Here, the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain. The difference between the speech signal to be processed and the late reverberation estimate is determined as the target speech signal. The accurate 2-norm weighting coefficients are determined through a neural network, thereby obtaining an accurate late reverberation estimate. The target speech signal is obtained by subtracting the late reverberation estimate from the speech signal to be processed, thereby achieving the effect of suppressing late reverberation.

[0055] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this disclosure, illustrate exemplary embodiments of the present disclosure and are used to explain the disclosure, but do not constitute an undue limitation of the disclosure. In the drawings:

[0057] Figure 1 This is a schematic diagram of a system architecture for suppressing late reverberation in an embodiment of this disclosure;

[0058] Figure 2 This is a schematic diagram of the process for suppressing late reverberation in an embodiment of this disclosure;

[0059] Figure 3 This is a schematic diagram of the process for determining late reverberation estimates based on filter coefficients in an embodiment of this disclosure;

[0060] Figure 4 This is a schematic diagram of the logic architecture of a device for suppressing late reverberation according to an embodiment of this disclosure;

[0061] Figure 5 This is a schematic diagram of the physical architecture of the smart terminal in an embodiment of this disclosure. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the technical solutions of this disclosure, and not all embodiments. Based on the embodiments recorded in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this disclosure.

[0063] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0064] The preferred embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0065] See Figure 1 As shown in this embodiment, the system includes at least one microphone array and a smart terminal. Figure 1 In Chinese, microphone arrays are a common type of voice device. During use, they generate raw voice signals, which inevitably contain noise and reverberation (including early and late reverberation, with late reverberation having a detrimental effect on playback). Therefore, directly playing the raw voice signal may result in poor playback quality or even complete playback failure. To address this, the raw voice signal is further input into a smart terminal for processing to suppress late reverberation and other noise, ensuring that the processed signal can be played normally.

[0066] In this embodiment of the disclosure, the method for suppressing late reverberation is mainly implemented on the smart terminal side, which will be described in detail below.

[0067] See Figure 2 As shown in the embodiments of this disclosure, the specific process for suppressing late reverberation is as follows:

[0068] Step 201: Denoise the original speech signal using a neural network, and determine the 2-norm weight coefficients based on the denoised signal.

[0069] It should be noted that the aforementioned original speech signal is the initial signal generated by a microphone array or other speech device in the time domain. To eliminate noise components in the original speech signal, a neural network is introduced in this embodiment to perform noise reduction processing on the original speech signal, and the result of the noise reduction processing is used as the final signal.

[0070] Here is a supplementary explanation of the training process of the above neural network: Speech data without noise and late-stage mixed estimates is pre-calculated manually using existing methods. This speech data is used as the output z(n) of the neural network. At the same time, the original speech data corresponding to the speech data is used as the input y(n) of the neural network. With the original speech data as input, the neural network is trained on the original speech data. That is, after running formula (1) through the neural network, the attribute function f(x) of the neural network is obtained. That is, the noise reduction processing of the original speech data is achieved through the attribute function f(x).

[0071] z(n)=f(y(n)) Formula (1)

[0072] Similarly, in subsequent applications, the original speech signal can also achieve noise reduction after being processed by this neural network. Noise reduction here can be understood as noise elimination, since noise cannot be completely eliminated; the goal is to control the noise within an acceptable range. Currently, commonly used neural network models include U-NET and FAS-NET.

[0073] The following details how to determine the 2-norm weighting coefficients based on the denoised signal:

[0074] (1) Convert the noise-reduced signal from the time domain to the frequency domain to obtain the reference signal.

[0075] In the field of speech signal processing, in order to facilitate feature extraction and recognition of speech signals, the speech signal is usually converted from the time domain to the frequency domain. After performing relevant processing on the speech signal in the frequency domain, the processed frequency domain signal is then converted back to the time domain for playback to meet user needs.

[0076] In this embodiment of the disclosure, the reference signal is the frequency domain signal corresponding to the result signal in the time domain, that is, the result signal after neural network noise reduction is converted from the time domain to the frequency domain to obtain the reference signal; similarly, the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain, that is, the original speech signal is converted from the time domain to the frequency domain to obtain the speech signal to be processed. Obviously, the speech signal to be processed still includes noise components.

[0077] Here, the noise-reduced signal is transformed from the time domain to the frequency domain to obtain the reference signal, specifically including:

[0078] 1) The noise-reduced signal is segmented in the time domain according to a specified step size to obtain multiple segmented signal results.

[0079] Since speech signals are continuous in time, in order to facilitate their processing, the noise-reduced signal is segmented in the time domain according to a specified step size during the implementation process. That is, a specified step size is preset in the time domain. Usually, the specified step size is a time function, and the duration of the time function is fixed.

[0080] According to the chronological order in the time domain, the result signal is truncated according to the duration corresponding to the specified step size. In this way, multiple continuous segments of the result signal can be obtained, that is, multiple segmented result signals.

[0081] 2) Multiply each segment result signal with a preset window to obtain multiple window signals.

[0082] Since time-domain truncation introduces spectral leakage, a preset window, i.e., a window function of preset length, is set in this embodiment to reduce the corresponding truncation effect. Each segmented result signal is then multiplied by the preset window to obtain multiple window signals. Compared with the segmented result signals, these multiple window signals retain more main lobe signals, thereby eliminating spectral leakage.

[0083] 3) Perform short-time Fourier transform on each window signal to obtain multiple sub-reference signals in the frequency domain.

[0084] During implementation, considering that it is speech processing, after obtaining multiple window signals, a short-time Fourier transform is performed on them, that is, the corresponding convolution operation is performed, thus obtaining multiple sub-reference signals in the frequency domain.

[0085] 4) Determine the frequency domain signal corresponding to the result signal based on each sub-reference signal, and use the frequency domain signal corresponding to the result signal as the reference signal.

[0086] After obtaining each sub-reference signal using the short-time Fourier transform, the sub-reference signals are combined in the frequency domain to obtain the frequency domain signal corresponding to the result signal, which is the reference signal.

[0087] Similarly, after converting the noise-free result signal to the frequency domain, the original speech signal containing noise is also converted to the frequency domain using the same conversion method. The following conversion process will not be repeated. The speech signal to be processed corresponding to the original speech signal is determined in the following way:

[0088] [1] The original speech signal is segmented in the time domain according to a specified step size to obtain multiple original segmented result signals.

[0089] [2] Multiply each original segment result signal with a preset window to obtain multiple original window signals.

[0090] [3] Perform short-time Fourier transform on each original window signal to obtain multiple sub-original reference signals in the frequency domain. Determine the frequency domain signal corresponding to the original speech signal based on each sub-original reference signal, and use the frequency domain signal corresponding to the original speech signal as the speech signal to be processed.

[0091] (2) Determine the power spectral density based on the amplitude value of the reference signal.

[0092] Power spectral density is a waveform used to characterize the power of a signal. It is specifically determined by the amplitude values ​​of a reference signal. Therefore, in practice, the corresponding power is calculated based on the amplitude values ​​of the reference signal in the frequency domain, and all calculated power values ​​are summarized into the power spectral density. This determined power spectral density is denoted as […].

[0093] (3) Determine the 2-norm weighting coefficients of the reference signal based on the average power spectral density, the number of channels of the reference signal, and the morphological parameters, where the number of channels and the morphological parameters are attributes of the reference signal.

[0094] To determine the 2-norm weighting coefficient w(n) of the reference signal, after obtaining the power spectral density, the average value of the power spectral density is further calculated. This average value is the mathematical result of summing and averaging the power spectral densities. Simultaneously, considering that the number of channels in the speech signal is uncertain—more channels correspond to a larger signal amplitude—it is necessary to extract relevant attributes from the reference signal, namely the number of channels M and the morphological parameter ε. Here, ε is a very small number that can be manually set.

[0095] Then, the 2-norm weighting coefficient of the reference signal is determined based on the average power spectral density, the number of channels of the reference signal, and the morphological parameters, as detailed in formula (2) below.

[0096]

[0097] After obtaining the 2-norm weighting coefficients of the reference signal, the 2-norm weighting coefficients and the speech signal to be processed are further input into the RLS weight prediction error multi-channel reverberation suppression algorithm, that is, the filter coefficients G(n) are calculated by using the existing algorithm of the RLS weight prediction error multi-channel reverberation suppression algorithm.

[0098] Step 202: Input the 2-norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain the filter coefficients, and determine the late reverberation estimate based on the filter coefficients. The speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain.

[0099] After obtaining the filter coefficients, further, based on the filter coefficients, determine the late reverberation estimate, see [reference needed]. Figure 3 As shown, it specifically includes:

[0100] Step 2021: Transpose the filter coefficients to obtain the transposed filter coefficients.

[0101] To facilitate subsequent calculations, the calculated filter coefficients G(n) are transposed during the implementation process to obtain the transposed filter coefficients G. H (n).

[0102] Step 2022: Determine a signal buffer block based on the delay time of the reference signal, wherein the signal buffer block includes the early reverberation of the reference signal.

[0103] Since the reference signal inevitably includes early reverberation, which in turn contains speech-related information, it is necessary to determine the delay time τ of the reference signal during implementation and to determine the signal buffer block based on the delay time τ. The signal buffer block Used to characterize the useful components in the reference signal.

[0104] Step 2023: Determine the late reverberation estimate based on the transposed filter coefficients and the signal buffer block.

[0105] During implementation, after determining the transposed filter coefficients and signal buffer block, the late reverberation estimate is calculated using the following formula (3).

[0106]

[0107] Step 203: The difference between the speech signal to be processed and the late reverberation estimate is determined as the target speech signal.

[0108] Since the late reverberation estimate is obtained through noise reduction using a neural network, the late reverberation estimate calculated on this basis is accurate. By subtracting the late reverberation estimate from the speech signal to be processed in the frequency domain, the influence of noise and late reverberation can be eliminated, and the above difference can be determined as the target speech signal.

[0109] In addition, since speech signals in the frequency domain cannot be played directly, after obtaining the target speech signal, it is converted to the time domain using inverse short-time Fourier transform to obtain a playable speech signal. This allows the speech signal to be played without noise and late reverberation, and the clarity of the played speech signal is higher than that of the original speech signal.

[0110] Based on the same inventive concept, see [reference] Figure 4As shown, this disclosure provides an apparatus for suppressing late reverberation, comprising:

[0111] The coefficient determination unit 401 is used to perform noise reduction processing on the original speech signal through a neural network, and determine the 2-norm weight coefficients based on the result signal after noise reduction processing.

[0112] Late reverberation determination unit 402 is used to input the 2-norm weight coefficients and the speech signal to be processed corresponding to the original speech signal into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients, and determine the late reverberation estimate based on the filter coefficients, wherein the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain.

[0113] The speech determination unit 403 is used to determine the difference between the speech signal to be processed and the late reverberation estimate as the target speech signal.

[0114] Optionally, the 2-norm weighting coefficients of the reference signal are determined based on the denoised result signal, and the coefficient determination unit 401 is used for:

[0115] The noise-reduced signal is converted from the time domain to the frequency domain to obtain the reference signal;

[0116] The power spectral density is determined based on the amplitude value of the reference signal;

[0117] The 2-norm weighting coefficients of the reference signal are determined based on the average power spectral density, the number of channels of the reference signal, and the morphological parameters, where the number of channels and the morphological parameters are attributes of the reference signal.

[0118] Optionally, the noise-reduced signal is converted from the time domain to the frequency domain to obtain a reference signal, and the late reverberation determination unit 402 is used for:

[0119] The noise-reduced signal is segmented in the time domain according to a specified step size to obtain multiple segmented signal results.

[0120] The segmented result signals are multiplied by a preset window to obtain multiple window signals;

[0121] Perform a short-time Fourier transform on each window signal to obtain multiple sub-reference signals in the frequency domain;

[0122] The frequency domain signal corresponding to the result signal is determined based on each sub-reference signal, and the frequency domain signal corresponding to the result signal is used as the reference signal.

[0123] Optionally, the speech signal to be processed corresponding to the original speech signal is determined in the following way:

[0124] The original speech signal is segmented in the time domain according to a specified step size to obtain multiple original segmented result signals;

[0125] The original segmented result signals are multiplied by the preset window to obtain multiple original window signals;

[0126] The original window signals are subjected to short-time Fourier transform to obtain multiple sub-original reference signals in the frequency domain. The frequency domain signal corresponding to the original speech signal is determined based on each sub-original reference signal, and the frequency domain signal corresponding to the original speech signal is used as the speech signal to be processed.

[0127] Optionally, the late reverberation estimate is determined based on the filter coefficients, and the late reverberation determination unit 402 is used for:

[0128] Transpose the filter coefficients to obtain the transposed filter coefficients;

[0129] Based on the delay time of the reference signal, a signal buffer block is determined, wherein the signal buffer block includes the early reverberation of the reference signal;

[0130] Based on the transposed filter coefficients and the signal buffer block, the late reverberation estimate is determined.

[0131] Based on the same inventive concept, see [reference] Figure 5 As shown, this embodiment of the present disclosure provides a smart terminal, including: a memory 501 for storing executable instructions; and a processor 502 for reading and executing the executable instructions stored in the memory, and executing any of the methods described in the first aspect above.

[0132] Based on the same inventive concept, embodiments of this disclosure provide a computer-readable storage medium that, when instructions in the storage medium are executed by a processor, enables the processor to perform the method described in any of the first aspects above.

[0133] In summary, the embodiments of this disclosure provide a method, apparatus, and storage medium for suppressing late reverberation. The method includes: performing noise reduction processing on the original speech signal using a neural network, determining the 2-norm weighting coefficients based on the noise reduction result signal, inputting the 2-norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal into an RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients, and determining a late reverberation estimate based on the filter coefficients. Here, the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain. The difference between the speech signal to be processed and the late reverberation estimate is determined as the target speech signal. The accurate 2-norm weighting coefficients are determined through a neural network, thereby obtaining an accurate late reverberation estimate. The target speech signal is obtained by subtracting the late reverberation estimate from the speech signal to be processed, thereby achieving the effect of suppressing late reverberation.

[0134] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program product systems. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product system implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program product systems according to this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A method of suppressing late reverberation, characterized by, The method includes: The original speech signal is denoised using a neural network. The denoised signal is then segmented in the time domain according to a specified step size to obtain multiple segmented result signals. Each segmented result signal is multiplied by a preset window to obtain multiple window signals. Each window signal is then subjected to a short-time Fourier transform to obtain multiple sub-reference signals in the frequency domain. The frequency domain signal corresponding to the result signal is determined based on each sub-reference signal, and this frequency domain signal is used as the reference signal. The power spectral density is determined based on the amplitude value of the reference signal. The 2-norm weighting coefficient of the reference signal is determined based on the average value of the power spectral density, the number of channels of the reference signal, and the morphological parameters, where the number of channels and the morphological parameters are attributes of the reference signal. The L2-norm weighting coefficients and the speech signal to be processed corresponding to the original speech signal are input into the RLS weight prediction error multi-channel reverberation suppression algorithm to obtain filter coefficients. The filter coefficients are then transposed to obtain transposed filter coefficients. Based on the delay time of the reference signal, a signal buffer block is determined, wherein the signal buffer block includes the early reverberation of the reference signal. Based on the transposed filter coefficients and the signal buffer block, the late reverberation estimate is determined, wherein the speech signal to be processed is the frequency domain signal corresponding to the original speech signal in the time domain. The difference between the speech signal to be processed and the late reverberation estimate is determined as the target speech signal.

2. The method of claim 1, wherein, The original speech signal corresponding to the speech signal to be processed is determined in the following way: The original speech signal is segmented in the time domain according to the specified step size to obtain multiple original segmented result signals; Each of the original segmented result signals is multiplied by the preset window to obtain multiple original window signals; The original window signals are subjected to short-time Fourier transform to obtain multiple sub-original reference signals in the frequency domain. The frequency domain signal corresponding to the original speech signal is determined based on each of the sub-original reference signals, and the frequency domain signal corresponding to the original speech signal is used as the speech signal to be processed.

3. An apparatus for suppressing late reverberation, characterized by include: The coefficient determination unit is used to perform noise reduction processing on the original speech signal through a neural network, and to segment the noise-reduced result signal in the time domain according to a specified step size to obtain multiple segmented result signals; to multiply each of the segmented result signals with a preset window to obtain multiple window signals; to perform short-time Fourier transform on each of the window signals to obtain multiple sub-reference signals corresponding to the frequency domain; to determine the frequency domain signal corresponding to the result signal based on each of the sub-reference signals, and to use the frequency domain signal corresponding to the result signal as the reference signal; Based on the amplitude value of the reference signal, the power spectral density is determined; based on the average value of the power spectral density, the number of channels of the reference signal, and the shape parameters, the 2-norm weighting coefficients of the reference signal are determined, wherein the number of channels and the shape parameters are attributes of the reference signal. The late reverberation determination unit is configured to input a 2-norm weight coefficient and a to-be-processed speech signal corresponding to an original speech signal into an RLS weight prediction error multi-channel reverberation suppression algorithm, obtain a filter coefficient, and transpose the filter coefficient to obtain a transposed filter coefficient; determine a signal buffer block based on a delay time of the reference signal, wherein the signal buffer block includes early reverberation of the reference signal; and determine the late reverberation estimation value based on the transposed filter coefficient and the signal buffer block, wherein the to-be-processed speech signal is a frequency domain signal corresponding to the original speech signal in a time domain. The speech determination unit is configured to determine a difference between the to-be-processed speech signal and the late reverberation estimation value as a target speech signal.

4. A smart terminal, characterized by The method comprises: a memory configured to store executable instructions; a processor configured to read and execute the executable instructions stored in the memory to implement the method according to any one of claims 1-2.

5. A computer readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor, the processor can execute the method according to any one of claims 1-2.

Images