Speech signal processing apparatus, method, electronic device, and sound amplification system

Abstract

This invention provides a speech signal processing apparatus, method, electronic device, and loudspeaker system, relating to the field of speech processing technology. The apparatus includes: a first echo cancellation unit, used to input a reference signal and a current speech signal acquired by a speech acquisition device into an adaptive filter to obtain a target residual signal; and a second echo cancellation unit, used to input the target residual signal and a far-end speech signal into a preset speech processing model to obtain a near-end speech processing signal for the current frame. This invention performs echo cancellation processing on the target residual signal using a speech processing model. Since the speech processing model is trained based on residual signal samples and far-end speech signal samples, and both the residual signal samples and the target residual signal include nonlinear echoes and other signals that the adaptive filter cannot completely eliminate, the speech processing model can eliminate the echoes of nonlinear components in the target residual signal, thereby improving the accuracy of speech signal processing.

Description

Technical Field

[0001] This invention relates to the field of speech processing technology, and in particular to a speech signal processing apparatus, method, electronic device, and amplification system. Background Technology

[0002] Echo refers to the signal from a distant speech signal emitted by a near-end speaker that is reflected by objects such as walls and then transmitted to a near-end microphone. The echo signal mixes with the near-end speech signal and is transmitted to the far-end speaker, causing the distant user to hear their own voice through the far-end speaker. To improve communication quality, echo cancellation technology is widely used in scenarios such as telephone calls and video conferencing. The most common echo cancellation method is through adaptive filters.

[0003] In related technologies, it is assumed that the near-end speech signal, which is used as the input signal, is uncorrelated with the far-end speech signal, which is used as the reference signal. The adaptive filter coefficients are iteratively updated with the goal of minimizing the correlation between the output signal of the adaptive filter and the far-end speech signal played by the speaker. The transmission path from the near-end speaker to the near-end microphone is modeled based on the updated filter coefficients. The echo signal is estimated using the far-end speech signal. The estimated echo signal is then subtracted from the speech signal collected by the near-end microphone, and the echo-removed speech signal is output.

[0004] However, among the aforementioned technologies, adaptive filters are linear operations and cannot remove echoes caused by nonlinear components such as speakers, acoustic channels, and microphones, thus reducing the accuracy of speech signal processing. Summary of the Invention

[0005] To address the problems existing in the prior art, embodiments of the present invention provide a speech signal processing apparatus, method, electronic device, and loudspeaker system.

[0006] This invention provides a speech signal processing device, comprising:

[0007] The first echo cancellation unit is used to input a reference signal and the current speech signal acquired by the speech acquisition device into an adaptive filter, and to process the current speech signal based on the reference signal through the adaptive filter to obtain the target residual signal; the reference signal includes the received far-end speech signal;

[0008] The second echo cancellation unit is used to input the target residual signal and the far-end speech signal into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0009] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0010] Furthermore, the reference signal also includes the previous frame of near-end speech processing signal output by the speech processing model.

[0011] Furthermore, the speech processing model includes a first feature extraction network, a second feature extraction network, a third feature extraction network, and a multi-layer artificial neural network;

[0012] The second echo cancellation unit is specifically used for:

[0013] The target residual signal is input into the first feature extraction network, and the target residual signal is mapped from the time domain to the transform domain through the first feature extraction network to obtain the first feature;

[0014] The remote speech signal is input into the second feature extraction network, and the second feature extraction network maps the remote speech signal from the time domain to the transform domain to obtain the second feature.

[0015] Both the first feature and the second feature are input into the multilayer artificial neural network. Based on the first feature and the second feature, the multilayer artificial neural network extracts the mask of the near-end speech signal of the current frame from the first feature.

[0016] The third feature of the current frame near-end speech signal in the transform domain is determined based on the mask and the first feature;

[0017] The third feature is input into the third feature extraction network, and the third feature is mapped from the transform domain to the time domain through the third feature extraction network to obtain the current frame near-end speech processing signal.

[0018] Furthermore, the speech processing model is obtained based on the following method:

[0019] The residual signal sample and the far-end speech signal sample are input into the initial network model to obtain the near-end speech processing sample output by the initial network model.

[0020] The loss function is determined based on the near-end speech processing samples and the desired signal; the desired signal includes near-end residual signal samples.

[0021] The model parameters of the initial network model are optimized based on the loss function until the convergence condition is met, thus obtaining the speech processing model.

[0022] Furthermore, the device also includes:

[0023] The sample acquisition unit is used to acquire near-end noisy signal samples and impulse response samples from the speech playback device to the speech acquisition device.

[0024] The sample delay unit is used to delay the near-end noisy signal sample for a preset time to obtain the target near-end noisy signal sample.

[0025] The echo sample determination unit is used to determine the near-end speech echo sample based on the target near-end noisy signal sample and the impulse response sample;

[0026] An input signal sample determination unit is used to determine an input signal sample based on the near-end speech echo sample and the near-end noisy signal sample;

[0027] A near-end residual signal sample determination unit is used to input the input signal sample and the target near-end noisy signal sample into the adaptive filter, and to process the input signal sample based on the target near-end noisy signal sample through the adaptive filter to obtain the near-end residual signal sample;

[0028] A residual signal sample determination unit is used to determine the residual signal sample based on the proximal residual signal sample.

[0029] Furthermore, the residual signal sample determination unit is specifically used for:

[0030] Based on the remote speech signal samples and impulse response samples, determine the remote speech echo samples;

[0031] The far-end speech echo sample and the far-end speech signal sample are input into the adaptive filter. The adaptive filter processes the far-end speech echo sample based on the far-end speech signal sample to obtain the far-end residual signal sample.

[0032] The residual signal sample is determined based on the near-end residual signal sample and the far-end residual signal sample.

[0033] Furthermore, the echo sample determination unit is specifically used for:

[0034] Based on the target near-end noisy signal sample and the impulse response sample, a reference near-end speech echo sample is determined. The reference near-end speech echo sample is delayed by the preset time to obtain a delayed near-end speech echo sample. The delayed near-end speech echo sample is used as a new target near-end noisy signal sample. The above steps are repeated until the number of delays reaches the preset number.

[0035] The near-end speech echo sample is determined based on the reference near-end speech echo sample obtained each time.

[0036] Furthermore, the first echo cancellation unit is specifically used for:

[0037] The current speech signal is processed by the adaptive filter based on the reference signal to obtain an output signal;

[0038] With the goal of minimizing the correlation between the output signal and the reference signal, the current impulse response of the adaptive filter is updated to obtain the target impulse response;

[0039] The target echo signal is determined based on the target impulse response and the reference signal;

[0040] The current speech signal is processed based on the target echo signal to obtain the target residual signal.

[0041] This invention provides a speech signal processing method, comprising:

[0042] The reference signal and the current speech signal acquired by the speech acquisition device are input into an adaptive filter. The adaptive filter processes the current speech signal based on the reference signal to obtain the target residual signal. The reference signal includes the received far-end speech signal.

[0043] The target residual signal and the far-end speech signal are input into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0044] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0045] The present invention also provides a loudspeaker system, comprising:

[0046] A voice acquisition device is used to acquire the current voice signal and input the current voice signal to a voice signal processing device;

[0047] The speech signal processing device is any of the speech signal processing devices described above.

[0048] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described speech signal processing methods.

[0049] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the speech signal processing method as described above.

[0050] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the speech signal processing method as described above.

[0051] The speech signal processing apparatus, method, electronic device, and amplification system provided by this invention input the far-end speech signal and the target residual signal output by an adaptive filter into a pre-trained speech processing model to obtain the near-end speech processing signal of the current frame for transmission output by the speech processing model. It can be seen that this invention further de-echoes the target residual signal output by the adaptive filter through the speech processing model. Since the speech processing model is trained based on residual signal samples and far-end speech signal samples, and both the residual signal samples and the target residual signal include nonlinear echoes and other signals that the adaptive filter cannot completely eliminate, the speech processing model can eliminate the echoes of nonlinear components in the target residual signal, thereby improving the accuracy of speech signal processing. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is one of the flowcharts illustrating the speech signal processing method provided in this embodiment of the invention;

[0054] Figure 2 This is a schematic diagram of a voice interaction scenario provided in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram illustrating the principle of the adaptive filter provided in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of the speech processing model provided in an embodiment of the present invention;

[0057] Figure 5 This is a second schematic flowchart of the speech signal processing method provided in the embodiments of the present invention;

[0058] Figure 6 This is the third flowchart of the speech signal processing method provided in this embodiment of the invention;

[0059] Figure 7 This is the fourth flowchart of the speech signal processing method provided in the embodiments of the present invention;

[0060] Figure 8This is the fifth flowchart of the speech signal processing method provided in the embodiments of the present invention;

[0061] Figure 9 This is a schematic diagram of the structure of the speech signal processing device provided in an embodiment of the present invention;

[0062] Figure 10 This is a schematic diagram of the physical structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0064] The following is combined with Figures 1-8 The speech signal processing method of the present invention is described.

[0065] Figure 1 This is one of the flowcharts illustrating the speech signal processing method provided in this embodiment of the invention, such as... Figure 1 As shown, the speech signal processing method includes the following steps:

[0066] Step 101: Input the reference signal and the current speech signal acquired by the speech acquisition device into the adaptive filter. The adaptive filter processes the current speech signal based on the reference signal to obtain the target residual signal.

[0067] The voice acquisition device includes a microphone, the reference signal includes the received far-end voice signal, the current voice signal includes the near-end voice signal (the voice spoken by the near-end user), environmental noise, and the target signal. The target signal may include the far-end voice echo signal and the near-end voice echo signal. The far-end voice echo signal is the far-end voice signal acquired by the voice acquisition device and played by the voice playback device, and the near-end voice echo signal is the near-end voice signal acquired by the voice acquisition device and played by the voice playback device. The voice playback device includes a speaker.

[0068] For example, Figure 2 This is a schematic diagram of a voice interaction scenario provided by an embodiment of the present invention, such as... Figure 2As shown, User A and User B are having a conference call. User A's voice is captured by voice acquisition device A1 on User A's side and transmitted via network by electronic device A2 on User A's side to electronic device B1 on User B's side. Electronic device B1 then plays User A's voice through voice playback device B2 on User B's side. At this time, User B is also speaking. When voice acquisition device B3 on User B's side acquires the voice signal, it will capture not only User B's voice but also User A's voice played by voice playback device B2. If the voice signal acquired by voice acquisition device B3 is transmitted to voice playback device A3 on User A's side without echo cancellation processing, User A will hear both User B's voice and User A's own voice played by voice playback device A3. This phenomenon is called echo. The remote voice signal in step 101 above can be understood as User A's voice transmitted to User B's side.

[0069] Specifically, the current speech signal is processed by the adaptive filter based on the reference signal to obtain an output signal; the current impulse response of the adaptive filter is updated with the goal of minimizing the correlation between the output signal and the reference signal to obtain a target impulse response; a target echo signal is determined based on the target impulse response and the reference signal; and the current speech signal is processed based on the target echo signal to obtain the target residual signal.

[0070] For example, when a reference signal and the current speech signal acquired by the speech acquisition device are obtained, the current speech signal acquired by the speech acquisition device is used as the input signal. Both the input signal and the reference signal are input into an adaptive filter. The adaptive filter filters the far-end speech echo signal and the near-end speech echo signal in the current speech signal based on the reference signal to obtain the output signal. Then, the correlation between the output signal and the reference signal is calculated through an adaptive algorithm. With the goal of minimizing the correlation, the current impulse response of the adaptive filter is updated. The transmission path from the speech playback device to the speech acquisition device is modeled to obtain the target impulse response. Then, the target impulse response and the reference signal are convolved to obtain the target echo signal. Finally, the target echo signal is subtracted from the current speech signal to obtain the target residual signal. The target residual signal includes the near-end speech signal of the current frame and other signals besides the near-end speech signal of the current frame, such as nonlinear echo signals.

[0071] It should be noted that the adaptive filter described above can be any suitable structure, such as a single-filter adaptive filter or a double-filter adaptive filter. The adaptive filter can also be a time-domain filter or a frequency-domain filter, and this invention does not limit it in this regard.

[0072] Step 102: Input the target residual signal and the far-end speech signal into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0073] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on the far-end speech signal sample and the residual signal sample.

[0074] For example, when the target residual signal is obtained, the target residual signal is used as the input signal of the speech processing model and the far-end speech signal is used as the reference signal of the speech processing model. The speech processing model further processes the other signals in the target residual signal, except for the near-end speech signal of the current frame, based on the far-end speech signal, and finally outputs the near-end speech processing signal of the current frame. This near-end speech processing signal of the current frame is the speech signal used to be transmitted to the other end.

[0075] The speech signal processing method provided by this invention inputs the far-end speech signal and the target residual signal output by the adaptive filter into a pre-trained speech processing model to obtain the near-end speech processing signal of the current frame for transmission output by the speech processing model. It can be seen that this invention further de-echoes the target residual signal output by the adaptive filter through the speech processing model. Since the speech processing model is trained based on residual signal samples and far-end speech signal samples, and both the residual signal samples and the target residual signal include nonlinear echoes and other signals that the adaptive filter cannot completely eliminate, the speech processing model can eliminate the echoes of nonlinear components in the target residual signal, thereby improving the accuracy of speech signal processing. Furthermore, the reference signals of the adaptive filter and the speech processing model in this invention do not require hardware analog-to-digital converters for processing; they are soft references, thereby reducing hardware costs and circuit board wiring complexity, and improving the flexibility of electronic device deployment.

[0076] Optionally, the reference signal may further include the previous frame of near-end speech processing signal output by the speech processing model.

[0077] For example, such as Figure 2 In the voice interaction scenario shown, the previous frame of near-end voice processing signal can be understood as the voice processing model processing the previous frame of voice signal collected by the voice acquisition device B3 on the user B side, and then obtaining the voice signal for transmission to the user A side.

[0078] For example, Figure 3 This is a schematic diagram of the principle of the adaptive filter provided in an embodiment of the present invention, as shown below. Figure 3As shown, when the far-end speech signal f(n) and the previous frame near-end speech processing signal u(n-1) are obtained, the far-end speech signal f(n) and the previous frame near-end speech processing signal u(n-1) can be superimposed as a reference signal r1(n). Then, the reference signal r1(n) and the current speech signal y(n) acquired by the speech acquisition device are input into the adaptive filter, where y(n) = x1(n) + s(n) + v(n) + x2(n), x1(n) represents the far-end speech echo signal, s(n) represents the near-end speech signal, v(n) represents the ambient noise signal, and x2(n) represents the near-end speech echo signal. The signal is then passed through the adaptive filter base... The current speech signal y(n) is processed using the reference signal r1(n) to obtain the target residual signal e(n). The target residual signal e(n) and the far-end speech signal f(n) are input into the speech processing model to obtain the current frame near-end speech processing signal u(n) output by the speech processing model. Since the near-end speech signal s(n) is amplified by the local speech playback device, then acquired by the speech acquisition device, then amplified by the local speech playback device, and then acquired by the speech acquisition device, this cycle will produce howling. Therefore, the purpose of adding the previous frame near-end speech processing signal u(n-1) to the reference signal r1(n) is to break this cycle and suppress howling.

[0079] Furthermore, when the reference signal includes the previous frame's near-end speech processing signal and the far-end speech signal, and the input signal includes the current speech signal, the adaptive filter, when updating the current impulse response, assigns a higher weight to the far-end speech signal in the reference signal that satisfies the assumption that the input signal and the reference signal are uncorrelated, thus ensuring the accuracy of the modeling. Conversely, when the power of the howling suppression part, which is strongly correlated with the input signal and the reference signal, is high, it assigns a lower weight to ensure that the adaptive filter can converge normally even without the far-end speech signal, relying solely on the near-end speech signal, thus ensuring the stability of the adaptive filter's operation.

[0080] The speech signal processing method provided in this invention adds a reference signal to the previous frame's near-end speech processing signal, and processes the current speech signal using an adaptive filter based on the reference signal, which includes the previous frame's near-end speech processing signal and the far-end speech signal. This achieves far-end echo processing and howling suppression. In other words, the echo processing and howling suppression of this invention model the same transmission path from the speech playback device to the speech acquisition device, and implement echo processing and howling suppression within the same adaptive filter.

[0081] Optionally, Figure 4 This is a schematic diagram of the speech processing model provided in an embodiment of the present invention, as shown below. Figure 4As shown, the speech processing model includes a first feature extraction network 401, a second feature extraction network 402, a third feature extraction network 403, and a multilayer artificial neural network 404. The input of the first feature extraction network 401 serves as the input of the speech processing model. The output of the first feature extraction network 401 is connected to the input of the second feature extraction network 402. The output of the second feature extraction network 402 is connected to the input of the multilayer artificial neural network 403. The output of the multilayer artificial neural network 403 is connected to the input of the third feature extraction network 404. The output of the third feature extraction network 404 serves as the output of the speech processing model. Each of the first feature extraction network 401, the second feature extraction network 402, and the third feature extraction network 403 can be composed of a single fully connected layer, a one-dimensional convolutional layer, or multiple fully connected layers.

[0082] Figure 5 This is a second schematic flowchart of the speech signal processing method provided in this embodiment of the invention, as shown below. Figure 5 As shown, step 102 above can be implemented through the following steps:

[0083] Step 1021: Input the target residual signal into the first feature extraction network, and map the target residual signal from the time domain to the transform domain through the first feature extraction network to obtain the first feature.

[0084] The first feature extraction network can be a fully connected layer or a one-dimensional convolutional layer.

[0085] For example, when the adaptive filter is a time-domain filter, the target residual signal is a time-domain signal; when the adaptive filter is a frequency-domain filter, the target residual signal is a frequency-domain signal. It is necessary to convert the frequency-domain target residual signal into a time-domain target residual signal, and then input the time-domain target residual signal into the first feature extraction network. The first feature extraction network maps the target residual signal from the time domain to the transform domain learned by the speech processing model to obtain the first feature corresponding to the target residual signal.

[0086] It should be noted that when obtaining the target residual signal in the time domain, the target residual signal in the time domain can also be segmented according to a preset length and a preset overlap to obtain multiple residual signal segments in the time domain corresponding to the target residual signal. Each time, one residual signal segment is input into the first feature extraction network. The preset length and preset overlap can be set based on actual needs. For example, the preset length is 256 sampling points and the preset overlap is 50%. This invention does not limit this.

[0087] Step 1022: Input the remote speech signal into the second feature extraction network, and use the second feature extraction network to map the remote speech signal from the time domain to the transform domain to obtain the second feature.

[0088] The second feature extraction network can be a fully connected layer or a one-dimensional convolutional layer.

[0089] For example, when the adaptive filter is a time-domain filter, the far-end speech signal is a time-domain signal; when the adaptive filter is a frequency-domain filter, the far-end speech signal is a frequency-domain signal. It is necessary to convert the frequency-domain far-end speech signal into a time-domain far-end speech signal, and then input the far-end speech signal into the second feature extraction network. The second feature extraction network maps the far-end speech signal from the time domain to the transform domain learned by the speech processing model to obtain the second feature corresponding to the far-end speech signal.

[0090] It should be noted that when obtaining the far-end speech signal in the time domain, the far-end speech signal in the time domain can also be segmented according to a preset length and a preset overlap to obtain multiple far-end speech signal segments in the time domain corresponding to the far-end speech signal. Each time, a far-end speech signal segment is input into the second feature extraction network.

[0091] Step 1023: Input both the first feature and the second feature into the multilayer artificial neural network, and extract the mask of the near-end speech signal of the current frame from the first feature based on the first feature and the second feature.

[0092] Among them, multi-layer artificial neural networks can be recurrent neural networks based on Long Short-Term Memory (LSTM) networks and gated recurrent units (GRU), or multi-layer convolutional neural networks.

[0093] For example, upon obtaining the first and second features, these features are input into a multilayer artificial neural network. The multilayer artificial neural network then extracts a mask for the near-end speech signal of the current frame based on the first and second features. The mask consists of a set of coefficients, each representing the weight multiplied at the corresponding point in the transform domain for converting the input data of each frame into near-end speech.

[0094] It should be noted that when obtaining the first and second features, the first feature can be normalized, and the second feature can also be normalized. The normalized first feature and the normalized second feature are then input into the multilayer artificial neural network. This can further improve the accuracy of the multilayer artificial neural network in extracting the mask of the near-end speech signal of the current frame, and further improve the accuracy of the speech processing model.

[0095] Step 1024: Determine the third feature of the near-end speech signal of the current frame in the transform domain based on the mask and the first feature.

[0096] For example, when the mask of the near-end speech signal of the current frame is obtained, the mask is multiplied by the first feature to obtain the third feature of the near-end speech signal of the current frame in the transform domain.

[0097] Step 1025: Input the third feature into the third feature extraction network, and map the third feature from the transform domain to the time domain through the third feature extraction network to obtain the current frame near-end speech processing signal.

[0098] The third feature extraction network can be a fully connected layer or a one-dimensional convolutional layer.

[0099] For example, when the third feature in the transform domain is obtained, the third feature is input into the third feature extraction network, which maps the third feature from the transform domain to the time domain to obtain the near-end speech processing signal of the current frame.

[0100] It should be noted that each time a residual signal segment is input into the first feature input layer, and each time a far-end speech signal segment is input into the second feature input layer, the resulting near-end speech processing signal of the current frame is also a near-end speech processing signal segment. When all near-end speech processing signal segments are obtained, they are combined based on the time order and preset overlap to obtain the near-end speech processing signal of the current frame.

[0101] The speech signal processing method provided in this invention extracts features from the target residual signal using a first feature extraction network to obtain a first feature layer; extracts features from the far-end speech signal using a second feature extraction network to obtain a second feature layer; and then obtains the final near-end speech processing signal of the current frame based on a multi-layer artificial neural network and a third feature extraction network. This method eliminates other signals in the target residual signal besides the near-end speech processing signal of the current frame. It utilizes the nonlinear processing capability of the speech processing model to process the nonlinear components in the echo. In addition, since the speech processing model has memory capability, it can store the reference signal corresponding to a previous period of time, and then process the signal with severe reverberation based on multiple reference signals, further improving the accuracy of speech signal processing.

[0102] Optionally, Figure 6 This is the third flowchart of the speech signal processing method provided in this embodiment of the invention, as shown below. Figure 6 As shown, the training steps of the speech processing model are as follows:

[0103] Step 601: Input the residual signal sample and the far-end speech signal sample into the initial network model to obtain the near-end speech processing sample output by the initial network model.

[0104] For example, before training the speech processing model, multiple far-end speech signal samples are first obtained to form a far-end speech signal dataset. The residual signal samples corresponding to each far-end speech signal sample are then obtained to form a residual signal sample dataset. An initial network model is then constructed. A far-end speech signal sample is randomly selected from the far-end speech signal dataset as a reference sample and input into the initial network model. The residual signal sample corresponding to the far-end speech signal sample is selected from the residual signal sample dataset as an input sample and input into the initial network model. The initial network model processes the signals in the residual signal samples other than the near-end speech processing samples based on the far-end speech signal samples to obtain the near-end speech processing samples output by the initial network model.

[0105] Step 602: Determine the loss function based on the near-end speech processing samples and the desired signal.

[0106] The desired signal includes a near-end residual signal sample.

[0107] Specifically, a first loss function and a second loss function can be constructed based on near-end speech processing samples and desired signals, and the weighted sum of the first loss function and the second loss function is determined as the loss function; wherein, the first loss function is mainly used to ensure the accuracy of model convergence, and the second loss function is mainly used to ensure the stability of model convergence.

[0108] The first loss function is represented by the following formula (1):

[0109]

[0110] Among them, Loss SNR Let denot be the first loss function, u(n) represent the near-end speech processing sample, and gt(n) represent the desired signal.

[0111] The second loss function is expressed by the following formula (2):

[0112]

[0113] Among them, Loss SmoothL1 Let x represent the second loss function, x = u(n) - gt(n).

[0114] It should be noted that the loss function can also be a loss function related to Perceptual Evaluation of Speech Quality (PESQ) or Short-Time Objective Intelligence (STOI) used to improve the listening experience, and this invention does not limit it to this.

[0115] Step 603: Optimize the model parameters of the initial network model based on the loss function until the convergence condition is met, and obtain the speech processing model.

[0116] For example, when the loss function is obtained, the model parameters of the initial network model are optimized based on the loss function, and the process is iterated continuously until the number of iterations reaches the preset number, at which point the convergence condition is determined, and the speech processing model is obtained.

[0117] The speech signal processing method provided in this embodiment of the invention trains an initial network model based on residual signal samples and far-end speech signal samples, and optimizes the model parameters of the initial network model based on a loss function, finally obtaining a trained speech processing model. This facilitates further processing of the target residual signal based on the speech processing model in the later stage, thereby improving the accuracy of speech signal processing.

[0118] Optionally, Figure 7 This is the fourth flowchart of the speech signal processing method provided in this embodiment of the invention, as shown below. Figure 7 As shown, prior to step 401 above, the speech signal processing method further includes the following steps:

[0119] Step 604: Obtain near-end noisy signal samples and impulse response samples from the voice playback device to the voice acquisition device.

[0120] Among them, the near-end noisy signal sample is the superposition of near-end speech signal sample and environmental noise sample. That is, near-end speech signal sample and environmental noise sample are collected in the actual application scenario, and the near-end noisy signal sample is obtained by superimposing the near-end speech signal sample and the near-end noisy signal sample. Multiple near-end noisy signal samples are composed of near-end noisy signal dataset. In addition, the real impulse response from the speech playback device to the speech acquisition device in the actual application scenario is collected, and the real impulse response is analyzed and synthesized to obtain impulse response sample. Multiple impulse response samples are composed of impulse response dataset.

[0121] Step 605: Delay the near-end noisy signal sample for a preset time to obtain the target near-end noisy signal sample.

[0122] The preset time range is slightly larger than the fluctuation range of the algorithm's processing time. The algorithm's processing time is the time from when the current speech signal is acquired by the speech acquisition device to when the speech processing model outputs the near-end speech processing signal. The purpose of setting the preset time range slightly larger than the fluctuation range of the algorithm's processing time is to make the processing time of the final trained speech processing model closer to the actual processing time of the algorithm.

[0123] For example, a random near-end noisy signal sample from the near-end noisy signal dataset is delayed for a preset time to obtain the near-end noisy signal sample corresponding to the time after the preset delay. The near-end noisy signal sample corresponding to the time after the preset delay is determined as the target near-end noisy signal sample.

[0124] Step 606: Based on the target near-end noisy signal sample and the impulse response sample, determine the near-end speech echo sample.

[0125] For example, the target near-end noisy signal sample and a random impulse response sample from the impulse response dataset are convolved to obtain the near-end speech echo sample after local amplification. Here, local amplification refers to playing the near-end speech signal sample collected by the speech acquisition device through a local speech playback device, and the near-end speech echo sample refers to the near-end speech signal sample being collected by the speech acquisition device after being played through the local speech playback device.

[0126] Step 607: Determine the input signal sample based on the near-end speech echo sample and the near-end noisy signal sample.

[0127] For example, the input signal sample is obtained by superimposing the near-end speech echo sample and the near-end noisy signal sample.

[0128] Step 608: Input the input signal sample and the target near-end noisy signal sample into the adaptive filter, and process the input signal sample based on the target near-end noisy signal sample through the adaptive filter to obtain the near-end residual signal sample.

[0129] For example, the target near-end noisy signal sample is used as a reference sample and input together with the input signal sample into an adaptive filter. The adaptive filter processes the near-end speech echo sample in the input signal sample based on the target near-end noisy signal sample to obtain the near-end residual signal sample.

[0130] Step 609: Determine the residual signal sample based on the near-end residual signal sample.

[0131] For example, when obtaining the near-end residual signal sample, the near-end residual signal sample is used as the residual signal sample input to the initial network model.

[0132] The speech signal processing method provided in this embodiment of the invention processes the near-end speech echo samples in the input signal samples using an adaptive filter to obtain near-end residual signal samples, thereby making the residual signal samples input to the initial network model more accurate.

[0133] Optionally, Figure 8 This is the fifth flowchart illustrating the speech signal processing method provided in this embodiment of the invention, as shown below. Figure 8 As shown, step 609 above can be implemented through the following steps:

[0134] Step 6091: Determine the far-end speech echo sample based on the far-end speech signal sample and the impulse response sample.

[0135] Among them, the remote voice echo sample refers to the voice acquisition device that acquires the remote voice signal sample after it has been played by the local voice playback device.

[0136] For example, the far-end speech signal sample and the impulse response sample are convolved to obtain the far-end speech echo sample.

[0137] Step 6092: Input the far-end speech echo sample and the far-end speech signal sample into the adaptive filter, and process the far-end speech echo sample based on the far-end speech signal sample through the adaptive filter to obtain the far-end residual signal sample.

[0138] For example, a far-end speech echo sample is input as an input signal into an adaptive filter, and a far-end speech signal sample is input as a reference signal into the adaptive filter. The far-end speech echo sample is then processed by the adaptive filter based on the far-end speech signal sample to obtain a far-end residual signal sample.

[0139] Step 6093: Determine the residual signal sample based on the near-end residual signal sample and the far-end residual signal sample.

[0140] For example, the near-end residual signal sample and the far-end residual signal sample are superimposed as the residual signal sample input to the initial network model.

[0141] The speech signal processing method provided in this embodiment of the invention processes the near-end speech echo samples in the input signal samples using an adaptive filter to obtain near-end residual signal samples, and processes the far-end speech echo samples using an adaptive filter to obtain far-end residual signal samples. The near-end residual signal samples and far-end residual signal samples are superimposed to obtain residual signal samples, which further improves the accuracy of residual signal samples input to the initial network model.

[0142] Optionally, step 606 above can be implemented in the following ways:

[0143] Based on the target near-end noisy signal sample and the impulse response sample, a reference near-end speech echo sample is determined. The reference near-end speech echo sample is delayed by the preset time to obtain a delayed near-end speech echo sample. The delayed near-end speech echo sample is used as a new target near-end noisy signal sample. The above steps are repeated until the number of delays reaches the preset number.

[0144] The near-end speech echo sample is determined based on the reference near-end speech echo sample obtained each time.

[0145] The preset number of attempts can be selected based on actual needs; for example, the preset number of attempts is 3.

[0146] For example, to avoid incomplete suppression of the near-end speech signal leading to repeated acquisition, the reference near-end speech echo sample is iterated a preset number of times, assuming the preset number is 3. The reference near-end speech echo sample is represented by near_noisy_echo0(n). Then, near_noisy_echo0(n) is delayed by a preset time and convolved with the impulse response sample to obtain near_noisy_echo1(n). Near_noisy_echo1(n) is then delayed by a preset time and convolved with the impulse response sample. The product is obtained by convolving near_noisy_echo2(n) with the impulse response sample after a preset delay. Then, near_noisy_echo2(n) is convolved with the impulse response sample to obtain near_noisy_echo3(n). At this time, the number of delays reaches the preset number. Then, near_noisy_echo0(n) + near_noisy_echo1(n) + near_noisy_echo2(n) + near_noisy_echo3(n) is determined as the near-end speech echo sample.

[0147] The speech signal processing method provided in this embodiment of the invention iterates the reference near-end speech echo sample a preset number of times to avoid incomplete suppression of signals other than near-end speech signals, which are then repeatedly collected by the speech acquisition device, thereby improving the accuracy of the near-end speech echo sample.

[0148] The speech signal processing apparatus provided by the present invention will be described below. The speech signal processing apparatus described below can be referred to in correspondence with the speech signal processing method described above.

[0149] Figure 9 This is a schematic diagram of the structure of the speech signal processing device provided in an embodiment of the present invention, as shown below. Figure 9 As shown, the speech signal processing device 900 includes a first echo cancellation unit 901 and a second echo cancellation unit 902; wherein:

[0150] The first echo cancellation unit 901 is used to input a reference signal and a current speech signal acquired by a speech acquisition device into an adaptive filter, and to process the current speech signal based on the reference signal through the adaptive filter to obtain a target residual signal; the reference signal includes the received far-end speech signal;

[0151] The second echo cancellation unit 902 is used to input the target residual signal and the far-end speech signal into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0152] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0153] The speech signal processing apparatus provided by this invention inputs a far-end speech signal and a target residual signal output by an adaptive filter into a pre-trained speech processing model to obtain a near-end speech processing signal for transmission in the current frame, output by the speech processing model. It can be seen that this invention further de-echoes the target residual signal output by the adaptive filter through the speech processing model. Since the speech processing model is trained based on residual signal samples and far-end speech signal samples, and both the residual signal samples and the target residual signal include nonlinear echoes and other signals that the adaptive filter cannot completely eliminate, the speech processing model can eliminate the echoes of nonlinear components in the target residual signal, thereby improving the accuracy of speech signal processing.

[0154] Based on any of the above embodiments, the reference signal further includes the previous frame of near-end speech processing signal output by the speech processing model.

[0155] Based on any of the above embodiments, the speech processing model includes a first feature extraction network, a second feature extraction network, a third feature extraction network, and a multilayer artificial neural network; the second echo cancellation unit 902 is specifically used for:

[0156] The target residual signal is input into the first feature extraction network, and the target residual signal is mapped from the time domain to the transform domain through the first feature extraction network to obtain the first feature;

[0157] The remote speech signal is input into the second feature extraction network, and the second feature extraction network maps the remote speech signal from the time domain to the transform domain to obtain the second feature.

[0158] Both the first feature and the second feature are input into the multilayer artificial neural network. Based on the first feature and the second feature, the multilayer artificial neural network extracts the mask of the near-end speech signal of the current frame from the first feature.

[0159] The third feature of the current frame near-end speech signal in the transform domain is determined based on the mask and the first feature;

[0160] The third feature is input into the third feature extraction network, and the third feature is mapped from the transform domain to the time domain through the third feature extraction network to obtain the current frame near-end speech processing signal.

[0161] Based on any of the above embodiments, the speech processing model is obtained in the following manner:

[0162] The residual signal sample and the far-end speech signal sample are input into the initial network model to obtain the near-end speech processing sample output by the initial network model.

[0163] The loss function is determined based on the near-end speech processing samples and the desired signal; the desired signal includes near-end residual signal samples.

[0164] The model parameters of the initial network model are optimized based on the loss function until the convergence condition is met, thus obtaining the speech processing model.

[0165] Based on any of the above embodiments, the speech signal processing device 900 further includes:

[0166] The sample acquisition unit is used to acquire near-end noisy signal samples and impulse response samples from the speech playback device to the speech acquisition device.

[0167] The sample delay unit is used to delay the near-end noisy signal sample for a preset time to obtain the target near-end noisy signal sample.

[0168] The echo sample determination unit is used to determine the near-end speech echo sample based on the target near-end noisy signal sample and the impulse response sample;

[0169] An input signal sample determination unit is used to determine an input signal sample based on the near-end speech echo sample and the near-end noisy signal sample;

[0170] A near-end residual signal sample determination unit is used to input the input signal sample and the target near-end noisy signal sample into the adaptive filter, and to process the input signal sample based on the target near-end noisy signal sample through the adaptive filter to obtain the near-end residual signal sample;

[0171] A residual signal sample determination unit is used to determine the residual signal sample based on the proximal residual signal sample.

[0172] Based on any of the above embodiments, the residual signal sample determination unit is specifically used for:

[0173] Based on the remote speech signal samples and impulse response samples, determine the remote speech echo samples;

[0174] The far-end speech echo sample and the far-end speech signal sample are input into the adaptive filter. The adaptive filter processes the far-end speech echo sample based on the far-end speech signal sample to obtain the far-end residual signal sample.

[0175] The residual signal sample is determined based on the near-end residual signal sample and the far-end residual signal sample.

[0176] Based on any of the above embodiments, the echo sample determination unit is specifically used for:

[0177] Based on the target near-end noisy signal sample and the impulse response sample, a reference near-end speech echo sample is determined. The reference near-end speech echo sample is delayed by the preset time to obtain a delayed near-end speech echo sample. The delayed near-end speech echo sample is used as a new target near-end noisy signal sample. The above steps are repeated until the number of delays reaches the preset number.

[0178] The near-end speech echo sample is determined based on the reference near-end speech echo sample obtained each time.

[0179] Based on any of the above embodiments, the first echo cancellation unit 901 is specifically used for:

[0180] The current speech signal is processed by the adaptive filter based on the reference signal to obtain an output signal;

[0181] With the goal of minimizing the correlation between the output signal and the reference signal, the current impulse response of the adaptive filter is updated to obtain the target impulse response;

[0182] The target echo signal is determined based on the target impulse response and the reference signal;

[0183] The current speech signal is processed based on the target echo signal to obtain the target residual signal.

[0184] This invention provides a public address system, including a voice acquisition device and a voice signal processing device; wherein:

[0185] A voice acquisition device is used to acquire the current voice signal and input the current voice signal to a voice signal processing device;

[0186] The speech signal processing device described in any of the above embodiments is used.

[0187] Furthermore, the public address system also includes a voice playback device for playing the current voice signal and / or a remote voice signal.

[0188] Figure 10 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of the present invention, such as... Figure 10 As shown, the electronic device 1000 may include a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute a speech signal processing method, which includes: inputting a reference signal and a current speech signal acquired by a speech acquisition device into an adaptive filter; processing the current speech signal based on the reference signal using the adaptive filter to obtain a target residual signal; the reference signal includes a received far-end speech signal.

[0189] The target residual signal and the far-end speech signal are input into the preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0190] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0191] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0192] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the speech signal processing method provided by the above methods, the method including: inputting a reference signal and a current speech signal acquired by a speech acquisition device into an adaptive filter, and processing the current speech signal based on the reference signal through the adaptive filter to obtain a target residual signal; the reference signal including a received far-end speech signal;

[0193] The target residual signal and the far-end speech signal are input into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0194] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0195] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the speech signal processing method provided by the above methods, the method comprising: inputting a reference signal and a current speech signal acquired by a speech acquisition device into an adaptive filter, and processing the current speech signal based on the reference signal through the adaptive filter to obtain a target residual signal; wherein the reference signal includes a received far-end speech signal;

[0196] The target residual signal and the far-end speech signal are input into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission.

[0197] The speech processing model is used to perform echo removal on the target residual signal; the speech processing model is trained based on far-end speech signal samples and residual signal samples.

[0198] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0199] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A speech signal processing device, characterized in that, include: The first echo cancellation unit is used to input a reference signal and the current speech signal acquired by the speech acquisition device into an adaptive filter, and to process the current speech signal based on the reference signal through the adaptive filter to obtain the target residual signal; the reference signal includes the received far-end speech signal; The second echo cancellation unit is used to input the target residual signal and the far-end speech signal into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission. The speech processing model is used to perform echo removal processing on the target residual signal; The speech processing model is obtained based on the following method: Residual signal samples and far-end speech signal samples are input into the initial network model to obtain near-end speech processing samples output by the initial network model. The residual signal samples are determined based on near-end residual signal samples, which are obtained by processing the input signal samples through the adaptive filter. The input signal samples are determined based on near-end speech echo samples and near-end noisy signal samples. The near-end speech echo samples are signals that are played by the near-end speech signal samples through the local speech playback device and then collected by the speech acquisition device. The near-end noisy signal samples are the superposition of near-end speech signal samples and environmental noise samples. The loss function is determined based on the near-end speech processing samples and the desired signal; the desired signal includes near-end residual signal samples. The model parameters of the initial network model are optimized based on the loss function until the convergence condition is met, thus obtaining the speech processing model.

2. The speech signal processing device according to claim 1, characterized in that, The reference signal also includes the previous frame of near-end speech processing signal output by the speech processing model.

3. The speech signal processing device according to claim 1, characterized in that, The speech processing model includes a first feature extraction network, a second feature extraction network, a third feature extraction network, and a multi-layer artificial neural network; The second echo cancellation unit is specifically used for: The target residual signal is input into the first feature extraction network, and the target residual signal is mapped from the time domain to the transform domain through the first feature extraction network to obtain the first feature; The remote speech signal is input into the second feature extraction network, and the second feature extraction network maps the remote speech signal from the time domain to the transform domain to obtain the second feature. Both the first feature and the second feature are input into the multilayer artificial neural network. Based on the first feature and the second feature, the multilayer artificial neural network extracts the mask of the near-end speech signal of the current frame from the first feature. The third feature of the current frame near-end speech signal in the transform domain is determined based on the mask and the first feature; The third feature is input into the third feature extraction network, and the third feature is mapped from the transform domain to the time domain through the third feature extraction network to obtain the current frame near-end speech processing signal.

4. The speech signal processing apparatus according to claim 1, characterized in that, The device further includes: The sample acquisition unit is used to acquire near-end noisy signal samples and impulse response samples from the speech playback device to the speech acquisition device. The sample delay unit is used to delay the near-end noisy signal sample for a preset time to obtain the target near-end noisy signal sample. The echo sample determination unit is used to determine the near-end speech echo sample based on the target near-end noisy signal sample and the impulse response sample; An input signal sample determination unit is used to determine an input signal sample based on the near-end speech echo sample and the near-end noisy signal sample; A near-end residual signal sample determination unit is used to input the input signal sample and the target near-end noisy signal sample into the adaptive filter, and to process the input signal sample based on the target near-end noisy signal sample through the adaptive filter to obtain the near-end residual signal sample; A residual signal sample determination unit is used to determine the residual signal sample based on the proximal residual signal sample.

5. The speech signal processing apparatus according to claim 4, characterized in that, The residual signal sample determination unit is specifically used for: Based on the far-end speech signal sample and the impulse response sample, the far-end speech echo sample is determined; The far-end speech echo sample and the far-end speech signal sample are input into the adaptive filter. The adaptive filter processes the far-end speech echo sample based on the far-end speech signal sample to obtain the far-end residual signal sample. The residual signal sample is determined based on the near-end residual signal sample and the far-end residual signal sample.

6. The speech signal processing apparatus according to claim 4, characterized in that, The echo sample determination unit is specifically used for: Based on the target near-end noisy signal sample and the impulse response sample, a reference near-end speech echo sample is determined. The reference near-end speech echo sample is delayed by the preset time to obtain a delayed near-end speech echo sample. The delayed near-end speech echo sample is used as a new target near-end noisy signal sample. The above steps are repeated until the number of delays reaches the preset number. The near-end speech echo sample is determined based on the reference near-end speech echo sample obtained each time.

7. The speech signal processing apparatus according to any one of claims 1-6, characterized in that, The first echo cancellation unit is specifically used for: The current speech signal is processed by the adaptive filter based on the reference signal to obtain an output signal; With the goal of minimizing the correlation between the output signal and the reference signal, the current impulse response of the adaptive filter is updated to obtain the target impulse response; The target echo signal is determined based on the target impulse response and the reference signal; The current speech signal is processed based on the target echo signal to obtain the target residual signal.

8. A speech signal processing method, characterized in that, include: The reference signal and the current speech signal acquired by the speech acquisition device are input into an adaptive filter. The adaptive filter processes the current speech signal based on the reference signal to obtain the target residual signal. The reference signal includes the received far-end speech signal. The target residual signal and the far-end speech signal are input into a preset speech processing model to obtain the current frame near-end speech processing signal output by the speech processing model for transmission. The speech processing model is used to perform echo removal processing on the target residual signal; The speech processing model is obtained based on the following method: Residual signal samples and far-end speech signal samples are input into the initial network model to obtain near-end speech processing samples output by the initial network model. The residual signal samples are determined based on near-end residual signal samples, which are obtained by processing the input signal samples through the adaptive filter. The input signal samples are determined based on near-end speech echo samples and near-end noisy signal samples. The near-end speech echo samples are signals that are played by the near-end speech signal samples through the local speech playback device and then collected by the speech acquisition device. The near-end noisy signal samples are the superposition of near-end speech signal samples and environmental noise samples. The loss function is determined based on the near-end speech processing samples and the desired signal; the desired signal includes near-end residual signal samples. The model parameters of the initial network model are optimized based on the loss function until the convergence condition is met, thus obtaining the speech processing model.

9. A loudspeaker system, characterized in that, include: A voice acquisition device is used to acquire the current voice signal and input the current voice signal to a voice signal processing device; The speech signal processing device is the speech signal processing device according to any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the speech signal processing method as described in claim 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the speech signal processing method as described in claim 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the speech signal processing method as described in claim 8.

Images