Noise reduction method for edge end real-time voice

Abstract

The invention belongs to the technical field of voice noise reduction. The invention provides an edge end real-time voice noise reduction method. According to the embodiment of the invention, pre-emphasis, framing and windowing processing is carried out on the input voice signal, so that the high-frequency component of the signal is improved and the distortion of the signal is reduced. And performing fast Fourier transform on the preprocessed input voice signal to convert a time domain signal into an original frequency domain signal. And calculating the Mel frequency cepstrum coefficient of the original frequency domain signal to extract the MFCC characteristic parameter of the speech. Noise estimation is performed on the MFCC characteristic parameters by using the lightweight separable convolutional neural network, so that the parameter quantity and the calculation quantity are reduced, and efficient noise estimation is realized. And performing adaptive iterative Wiener filtering by using the generated noise estimation, and gradually eliminating noise in the voice signal to obtain a filtered frequency domain signal. And inverse fast Fourier transform is carried out on the filtered frequency domain signal, and a time domain signal is recovered. And outputting a time domain signal which is a voice signal after noise reduction.

Description

Technical Field

[0001] The disclosed embodiments relate to the technical field of speech noise reduction, and in particular to a method for noise reduction of real-time speech at an edge end. Background Art

[0002] With the rapid development of information technology, voice communication and speech recognition technologies have been widely used in various fields. In particular, in scenarios such as smartphones, smart assistants, and remote conferences, clear voice signal transmission and processing are crucial. However, in practical applications, voice signals are often interfered with by various noises, which can come from the environment (such as traffic noise, crowd noise, etc.) or the equipment itself (such as poor microphone quality, unstable signal transmission, etc.). These noises can significantly reduce the clarity and intelligibility of voice signals, and even lead to information transmission errors.

[0003] At present, speech noise reduction is mainly divided into traditional filter noise reduction and deep learning-based noise reduction. The traditional filter noise reduction method has a small amount of calculation and is suitable for offline real-time noise reduction scenarios at the edge, but the noise reduction effect is poor, especially it is difficult to filter out non-stationary noise; while the deep learning-based noise reduction solution has excellent noise reduction effect, but the model calculation is huge and difficult to deploy and use in edge scenarios with limited power consumption and computing resources.

[0004] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0006] The purpose of the embodiments of the present disclosure is to provide a method for reducing noise of real-time speech at an edge, thereby overcoming one or more problems caused by limitations and defects of related technologies at least to a certain extent.

[0007] According to an embodiment of the present disclosure, a method for reducing noise of real-time speech at an edge is provided, the method comprising: Acquire an input voice signal, and preprocess the input voice signal to obtain a plurality of frames of voice signals; For one frame of the speech signal, performing a fast Fourier transform on the speech signal to obtain an original frequency domain signal; Calculating the Mel-frequency cepstral coefficients of the original frequency domain signal to extract the MFCC feature parameters of the input speech signal; Inputting the MFCC feature parameters into a separable convolutional neural network that has been processed by INT8 fixed-point quantization to obtain a noise estimation result; wherein the separable convolutional neural network includes an input layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer, a fully connected layer and an output layer in series, and the noise estimation result includes a speech frame and a non-speech frame; Using a Wiener filter to perform adaptive iterative Wiener filtering on the speech frame and the non-speech frame respectively to obtain a filtered frequency domain signal; Performing an inverse fast Fourier transform on the filtered frequency domain signal to obtain a filtered time domain signal; All the speech signals are traversed to obtain the filtered time domain signals corresponding to all the speech signals, and all the filtered time domain signals are combined to obtain a noise-reduced speech signal.

[0008] Furthermore, the step of obtaining an input voice signal and preprocessing the input voice signal includes: Pre-emphasize the input speech signal using a pre-emphasis filter; Dividing the pre-emphasized input voice signal into a plurality of frame voice signals; Each frame of the speech signal is windowed using a Hamming window to reduce spectrum leakage.

[0009] Further, the expression of the pre-emphasis filter is:

[0010] in, The original input signal is The value of the sampling point, The original input signal is The value of the sampling point, The original input signal after pre-emphasis is The value of the sampling point, is the pre-emphasis coefficient; The Hamming window expression is:

[0011] in, is the Hamming window function in The value of the sampling point, is the number of sampling points per frame, is the index of the current sampling point, ranging from ; Each frame of the speech signal The windowed expression is:

[0012] in, For the speech signal The value of the sampling point; The windowed speech signal is The value of the sampling point.

[0013] Further, the step of calculating the Mel-frequency cepstral coefficients of the original frequency domain signal to extract the MFCC feature parameters of the input speech signal includes: The frequency of the original frequency domain signal is converted into Convert to Mel frequency ; Wherein, the Mel filter bank includes a plurality of Mel filters; According to the Mel frequency and the original frequency domain signal, calculating the energy of each Mel filter; The energy of all the Mel filters is logarithmized and discrete cosine transformed to obtain MFCC feature parameters.

[0014] Furthermore, the expression of the Mel frequency is:

[0015] The calculation expression of the energy of the Mel filter is:

[0016] in, is the frequency point, is the original frequency domain signal, is the original frequency domain signal in The power spectral density at the frequency point, For the The Mel filter Frequency response of the frequency point, is the starting frequency of the Mel filter, is the end frequency of the Mel filter; The expression of the discrete cosine transform is:

[0017] in, For the MFCC feature parameters, is the number of Mel filters, For the The energy of a Mel filter.

[0018] Furthermore, the step of inputting the MFCC feature parameters into a separable convolutional neural network processed by INT8 fixed-point quantization to obtain a noise estimation result includes: Inputting the MFCC feature parameters into the input layer, the first depth-wise separable convolutional layer and the second depth-wise separable convolutional layer in sequence, flattening the output of the second depth-wise separable convolutional layer and inputting it into the fully connected layer, to generate the noise estimation result; The output layer outputs the noise estimation result; wherein, if the VAD signal of the noise estimation result is higher than a preset threshold, the speech signal of the current frame is the non-speech frame; if the VAD signal of the noise estimation result is less than or equal to the threshold, the speech signal of the current frame is the speech frame.

[0019] Furthermore, the first depthwise separable convolution layer includes a first depthwise convolution and a first pointwise convolution, and the second depthwise separable convolution layer includes a second depthwise convolution and a second pointwise convolution.

[0020] Further, the step of using a Wiener filter to perform adaptive iterative Wiener filtering on the speech frame and the non-speech frame to obtain a filtered frequency domain signal includes: The average power spectrum of the silent segment is used as the initial noise power spectrum estimation, and the preset constant is used as the speech power spectrum; If the speech signal in the current frame is the non-speech frame, the noise power spectrum estimation is updated:

[0021] in, The current frame The noise power spectrum estimation of frequency points, is the updated current frame The noise power spectrum estimation of frequency points, The current frame The power spectral density of the frequency point, is the smoothing coefficient; If the speech signal in the current frame is the speech frame, maintaining the noise power spectrum estimation unchanged; Calculate the gain of the Wiener filter based on the updated noise power spectrum estimate and speech power spectrum :

[0022] The original frequency domain signal of the current frame is subjected to adaptive iterative Wiener filtering using the Wiener filter to obtain a filtered frequency domain signal; wherein the expression of the adaptive iterative Wiener filtering is:

[0023] in, is the filtered frequency domain signal, is the original frequency domain signal.

[0024] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects: In the embodiment of the present disclosure, through the above-mentioned edge end real-time speech denoising method, on the one hand, the input speech signal is pre-emphasized, framed and windowed to improve the high-frequency components of the signal and reduce the distortion of the signal. The pre-processed input speech signal is subjected to a fast Fourier transform to convert the time domain signal into the original frequency domain signal. The Mel frequency cepstrum coefficients of the original frequency domain signal are calculated to extract the MFCC feature parameters of the speech. The MFCC feature parameters are estimated by using a lightweight separable convolutional neural network, which not only reduces the number of parameters and the amount of calculation, but also realizes efficient noise estimation. Using the generated noise estimate, adaptive iterative Wiener filtering is performed to gradually eliminate the noise in the speech signal to obtain a filtered frequency domain signal. The filtered frequency domain signal is subjected to an inverse fast Fourier transform to restore the time domain signal. The output time domain signal is the denoised speech signal. On the other hand, the method can perform offline real-time denoising in edge end scenarios with limited power consumption and small computing resources, and the denoising effect is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0026] Figure 1 A step diagram showing a method for reducing noise of real-time speech at an edge terminal in an exemplary embodiment of the present disclosure is shown; Figure 2 A schematic diagram of a minus 20db synthesized audio of a noisy speech and an audio after noise reduction in an exemplary embodiment of the present disclosure is shown; Figure 3 A schematic diagram of an audio recording of a call with ambient noise and its noise-reduced audio in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0027] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0028] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0029] This example implementation provides a method for reducing noise in real-time speech at an edge. Figure 1 As shown in , the edge end real-time speech noise reduction method may include: step S101 to step S107.

[0030] Step S101: obtaining an input speech signal, and preprocessing the input speech signal to obtain a plurality of frames of speech signals; Step S102: for a frame of the speech signal, perform a fast Fourier transform on the speech signal to obtain an original frequency domain signal; Step S103: Calculating the Mel-frequency cepstral coefficients of the original frequency domain signal to extract MFCC feature parameters of the input speech signal; Step S104: inputting the MFCC feature parameters into a separable convolutional neural network that has been processed by INT8 fixed-point quantization to obtain a noise estimation result; wherein the separable convolutional neural network includes an input layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer, a fully connected layer and an output layer in series, and the noise estimation result includes a speech frame and a non-speech frame; Step S105: using a Wiener filter to perform adaptive iterative Wiener filtering on the speech frame and the non-speech frame respectively to obtain a filtered frequency domain signal; Step S106: performing an inverse fast Fourier transform on the filtered frequency domain signal to obtain a filtered time domain signal; Step S107: traverse all the speech signals to obtain the filtered time domain signals corresponding to all the speech signals, and merge all the filtered time domain signals to obtain a noise-reduced speech signal.

[0031] Through the above-mentioned edge-end real-time speech noise reduction method, on the one hand, the input speech signal is pre-emphasized, framed and windowed to increase the high-frequency components of the signal and reduce the distortion of the signal. The preprocessed input speech signal is subjected to a fast Fourier transform to convert the time domain signal into the original frequency domain signal. The Mel frequency cepstrum coefficients of the original frequency domain signal are calculated to extract the MFCC feature parameters of the speech. The MFCC feature parameters are estimated by using a lightweight separable convolutional neural network, which not only reduces the number of parameters and the amount of calculation, but also realizes efficient noise estimation. Using the generated noise estimate, adaptive iterative Wiener filtering is performed to gradually eliminate the noise in the speech signal to obtain a filtered frequency domain signal. The filtered frequency domain signal is subjected to an inverse fast Fourier transform to restore the time domain signal. The output time domain signal is the denoised speech signal. On the other hand, this method can perform offline real-time noise reduction in edge scenarios with limited power consumption and small computing resources, and the noise reduction effect is better.

[0032] Next, we will refer to Figures 1 to 3 Each step of the above-mentioned edge-end real-time speech noise reduction method in this example implementation is described in more detail.

[0033] In step S101, an input speech signal is acquired and preprocessed to obtain a plurality of frames of speech signals.

[0034] Specifically, the input speech signal is pre-emphasized by using a pre-emphasis filter; the pre-emphasized input speech signal is divided into a plurality of frames of speech signals; and each frame of speech signal is windowed by using a Hamming window to reduce spectrum leakage.

[0035] In one embodiment, the input voice signal is pre-emphasized to enhance the high frequency components. Applying the pre-emphasis filter, the formula is:

[0036] in, The original input signal is The value of the sampling point, The original input signal is The value of the sampling point, The original input signal after pre-emphasis is The value of the sampling point, is the pre-emphasis coefficient, which is used to control the degree of enhancement of high-frequency components. For example, The value is 0.95.

[0037] The pre-emphasized signal is divided into several frames, each with a length of 10 milliseconds and a frame shift of 5 milliseconds. A Hamming window is applied to each frame signal to reduce spectrum leakage. The windowing formula is as follows:

[0038] in, is the Hamming window function in The value of the sampling point; is the number of sampling points per frame, i.e., the frame length; is the index of the current sampling point, ranging from .

[0039] When the Hamming window is applied, each frame of speech signal The windowed version is:

[0040] in, For the speech signal The value of the sampling point; The windowed speech signal is The value of the sampling point.

[0041] In step S102, for one frame of the speech signal, a fast Fourier transform is performed on the speech signal to obtain an original frequency domain signal.

[0042] Specifically, a fast Fourier transform (FFT) is performed on each frame of the windowed signal to convert the time domain signal into a frequency domain signal to obtain a complex spectrum.

[0043] In step S103, the Mel-frequency cepstral coefficients of the original frequency domain signal are calculated to extract the MFCC feature parameters of the input speech signal.

[0044] Specifically, the frequency of the original frequency domain signal is converted to Convert to Mel frequency ; Among them, the Mel filter bank includes several Mel filters; according to the Mel frequency and the original frequency domain signal, calculate the energy of each Mel filter; take the logarithm of the energy of all Mel filters and perform discrete cosine transform to obtain MFCC feature parameters.

[0045] In one embodiment, the Mel-Frequency Cepstral Coefficients (MFCC) of the frequency domain signal are calculated. First, the spectrum is passed through the Mel filter bank and the energy of each filter is calculated. Then, the logarithm of the filter energy is taken and a discrete cosine transform (DCT) is performed to obtain the MFCC features. The specific process is as follows: First, a Mel filter bank is applied (for example, 40 filters are used in this embodiment): the frequency Convert to Mel frequency , the Mel frequency formula is as follows:

[0046] For each filter, calculate its output energy :

[0047] in, is the frequency point, is the original frequency domain signal, is the original frequency domain signal in The power spectral density at the frequency point, For the The Mel filter Frequency response of the frequency point, is the starting frequency of the Mel filter, is the end frequency of the Mel filter; The expression of discrete cosine transform is:

[0048] in, For the MFCC feature parameters (for example, in this embodiment, i =10), is the number of Mel filters (for example, in this embodiment, M=40), For the The energy of a Mel filter.

[0049] In step S104, the MFCC feature parameters are input into a separable convolutional neural network that has been processed by INT8 fixed-point quantization to obtain a noise estimation result; wherein the separable convolutional neural network includes a serial input layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer, a fully connected layer and an output layer, and the noise estimation result includes speech frames and non-speech frames.

[0050] More specifically, the MFCC feature parameters are sequentially input into the input layer, the first depth-separable convolutional layer, and the second depth-separable convolutional layer, and the output of the second depth-separable convolutional layer is flattened and input into the fully connected layer to generate a noise estimation result; The output layer outputs the noise estimation result; if the VAD signal of the noise estimation result is higher than the preset threshold, the current frame speech signal is a non-speech frame; if the VAD signal of the noise estimation result is less than or equal to the threshold, the current frame speech signal is a speech frame.

[0051] In addition, the first depthwise separable convolution layer includes a first depthwise convolution and a first pointwise convolution, and the second depthwise separable convolution layer includes a second depthwise convolution and a second pointwise convolution.

[0052] In one embodiment, a lightweight separable convolutional neural network (i.e., separable convolutional neural network) is constructed, and the MFCC features (i.e., MFCC feature parameters) are input into the designed lightweight separable convolutional neural network. The network consists of several depth-separable convolutional layers and performs fixed-point quantization processing, which can quickly and efficiently obtain noise estimation results. The specific steps are as follows: (1) Input features: From the previous steps, the Mel-frequency cepstral coefficients (MFCC) of each frame signal are calculated. These MFCC features will serve as the input of the neural network.

[0053] (2) Design of lightweight separable convolutional neural network 1) Network structure: Input layer: Input MFCC two-dimensional feature vector (number of single-frame samples × number of features).

[0054] Depthwise separable convolution layer (2 layers): Use depthwise separable convolution to reduce the number of parameters and computation. Each convolution layer is divided into depthwise convolution (convolution kernel 10×2) and pointwise convolution (convolution kernel 1×1), with 32 channels.

[0055] Activation function: ReLU activation function is applied after each separable convolutional layer.

[0056] Fully connected layer: The output of the convolutional layer is flattened and input into the fully connected layer to generate the noise estimation result.

[0057] 2) Network training Data preparation: Mark noisy speech and clean speech as 1, mark pure noise data as 0, and use 15 hours of speech data for training.

[0058] Loss function: Use the mean square error loss function.

[0059] Perceptual quantization training: In order to obtain a lightweight model, perceptual quantization training is used on the network structure to obtain a lightweight network after fixed-point quantization to reduce the amount of network calculation.

[0060] In a specific embodiment, the output of each layer of the neural network is: Input layer: 160×10 First depthwise separable convolution output: 151×32.

[0061] Second depthwise separable convolution output: 151×32.

[0062] After ReLU activation: 151×32.

[0063] Fully connected layer: flattened to 4832.

[0064] In step S105, adaptive iterative Wiener filtering is performed on the speech frame and the non-speech frame using a Wiener filter to obtain a filtered frequency domain signal.

[0065] Specifically, the key role of using a neural network (ie, a separable convolutional neural network) for noise estimation is to determine whether there is a speech signal in the current frame. In the adaptive iterative Wiener filtering step, it is determined whether to update the parameters of the Wiener filter based on the result of the speech activity detection.

[0066] (1) Voice Activity Detection (VAD) In the noise estimation step, the lightweight separable convolutional neural network is used to indicate whether the current frame contains speech. A threshold is set (the threshold is set to 0.8 in this embodiment), and if the VAD signal output by the network is higher than the threshold, the current frame is determined to be a non-speech frame.

[0067] (2) Wiener filter initialization At the beginning of the process, the average power spectrum of the silence segment is used as the initial noise power spectrum estimate, and a small constant value (0.0001) is used as the initial speech power spectrum.

[0068] (3) Iterative update Non-speech frame processing: Noise power spectrum update If the current frame is determined to be a non-speech frame, the noise power spectrum estimate is updated:

[0069] in, The current frame The noise power spectrum estimation of frequency points, is the updated current frame The noise power spectrum estimation of frequency points, The current frame The power spectral density of the frequency point, is the smoothing coefficient. In this embodiment, is 0.998.

[0070] Speech frame processing: maintaining the noise power spectrum unchanged If the current frame is determined to be a speech frame, the noise power spectrum estimation is not updated.

[0071] (4) Wiener filter calculation Calculate the Wiener filter gain based on the updated noise and speech power spectra :

[0072] Then apply the Wiener filter to the frequency domain signal of the current frame for filtering:

[0073] in, is the filtered frequency domain signal, is the original frequency domain signal.

[0074] In step S106 and step S107, the filtered frequency domain signal is subjected to inverse fast Fourier transform to obtain a filtered time domain signal; all speech signals are traversed to obtain the filtered time domain signals corresponding to all speech signals, and all filtered time domain signals are merged to obtain a noise-reduced speech signal. Specifically, the filtered frequency domain signal is subjected to an inverse fast Fourier transform (IFFT) to restore the time domain signal, and the time domain signals of all frames are combined to obtain the final noise-reduced speech signal.

[0075] In a specific embodiment, the edge real-time speech noise reduction method proposed in this application (i.e., RTC method) is compared with the traditional noise reduction methods (specifically including RNNoise, DeepFilterNet and S-DCCRN). Running on stm32F407 (main frequency 168MHz), the results obtained are shown in Table 1. The power consumption of the edge real-time speech noise reduction method proposed in this application is only 0.8W, and it takes only 4ms to process a frame of 10ms audio. The noise suppression in a strong noise environment can reach more than 25db. It can be seen that compared with traditional methods and pure deep learning solutions, this method has fewer parameters and calculations, and has better noise reduction effect.

[0076] Table 1 Comparison of various noise reduction methods

[0077] like Figure 2 The figure shows a schematic diagram of the synthesized audio of noisy speech minus 20db and the audio after noise reduction.

[0078] like Figure 3 The figure shows a schematic diagram of the audio of a call recording in ambient noise and its noise reduction.

[0079] Through the above-mentioned edge-end real-time speech noise reduction method, on the one hand, the input speech signal is pre-emphasized, framed and windowed to increase the high-frequency components of the signal and reduce the distortion of the signal. The preprocessed input speech signal is subjected to a fast Fourier transform to convert the time domain signal into the original frequency domain signal. The Mel frequency cepstrum coefficients of the original frequency domain signal are calculated to extract the MFCC feature parameters of the speech. The MFCC feature parameters are estimated by using a lightweight separable convolutional neural network, which not only reduces the number of parameters and the amount of calculation, but also realizes efficient noise estimation. Using the generated noise estimate, adaptive iterative Wiener filtering is performed to gradually eliminate the noise in the speech signal to obtain a filtered frequency domain signal. The filtered frequency domain signal is subjected to an inverse fast Fourier transform to restore the time domain signal. The output time domain signal is the denoised speech signal. On the other hand, this method can perform offline real-time noise reduction in edge scenarios with limited power consumption and small computing resources, and the noise reduction effect is better.

[0080] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0081] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present disclosure can be understood according to specific circumstances.

[0082] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0083] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A method for reducing noise of real-time speech at an edge, characterized in that: The method includes: Acquire an input voice signal, and preprocess the input voice signal to obtain a plurality of frames of voice signals; For one frame of the speech signal, performing a fast Fourier transform on the speech signal to obtain an original frequency domain signal; Calculating the Mel-frequency cepstral coefficients of the original frequency domain signal to extract the MFCC feature parameters of the input speech signal; Inputting the MFCC feature parameters into a separable convolutional neural network that has been processed by INT8 fixed-point quantization to obtain a noise estimation result; wherein the separable convolutional neural network includes an input layer, a first depth-separable convolutional layer, a second depth-separable convolutional layer, a fully connected layer, and an output layer in series, and the noise estimation result includes a speech frame and a non-speech frame; Using a Wiener filter to perform adaptive iterative Wiener filtering on the speech frame and the non-speech frame respectively to obtain a filtered frequency domain signal; Performing an inverse fast Fourier transform on the filtered frequency domain signal to obtain a filtered time domain signal; All the speech signals are traversed to obtain the filtered time domain signals corresponding to all the speech signals, and all the filtered time domain signals are combined to obtain a noise-reduced speech signal.

2. The method for reducing noise of real-time speech at the edge according to claim 1, characterized in that: The step of obtaining an input voice signal and preprocessing the input voice signal includes: Pre-emphasize the input speech signal using a pre-emphasis filter; Dividing the pre-emphasized input voice signal into a plurality of frame voice signals; Each frame of the speech signal is windowed using a Hamming window to reduce spectrum leakage.

3. The method for reducing noise of real-time speech at the edge according to claim 2, characterized in that: The expression of the pre-emphasis filter is: in, The original input signal is The value of the sampling point, The original input signal is The value of the sampling point, The original input signal after pre-emphasis is The value of the sampling point, is the pre-emphasis coefficient; The Hamming window expression is: in, is the Hamming window function in The value of the sampling point, is the number of sampling points per frame, is the index of the current sampling point, ranging from ; Each frame of the speech signal The windowed expression is: in, For the speech signal The value of the sampling point; The windowed speech signal is The value of the sampling point.

4. The method for reducing noise of real-time speech at the edge according to claim 3, characterized in that: The step of calculating the Mel-frequency cepstral coefficients of the original frequency domain signal to extract the MFCC feature parameters of the input speech signal includes: The frequency of the original frequency domain signal is converted into Convert to Mel frequency ; Wherein, the Mel filter bank includes a plurality of Mel filters; According to the Mel frequency and the original frequency domain signal, calculating the energy of each Mel filter; The energy of all the Mel filters is logarithmized and discrete cosine transformed to obtain MFCC feature parameters.

5. The method for reducing noise of real-time speech at the edge according to claim 4, characterized in that: The expression of the Mel frequency is: The calculation expression of the energy of the Mel filter is: in, is the frequency point, is the original frequency domain signal, is the original frequency domain signal in The power spectral density at the frequency point, For the The Mel filter Frequency response of the frequency point, is the starting frequency of the Mel filter, is the end frequency of the Mel filter; The expression of the discrete cosine transform is: in, For the MFCC feature parameters, is the number of Mel filters, For the The energy of a Mel filter.

6. The method for reducing noise of real-time speech at the edge according to claim 5, characterized in that: The step of inputting the MFCC feature parameters into a separable convolutional neural network processed by INT8 fixed-point quantization to obtain a noise estimation result includes: Inputting the MFCC feature parameters into the input layer, the first depth-wise separable convolutional layer and the second depth-wise separable convolutional layer in sequence, flattening the output of the second depth-wise separable convolutional layer and inputting it into the fully connected layer, to generate the noise estimation result; The output layer outputs the noise estimation result; wherein, if the VAD signal of the noise estimation result is higher than a preset threshold, the speech signal of the current frame is the non-speech frame; if the VAD signal of the noise estimation result is less than or equal to the threshold, the speech signal of the current frame is the speech frame.

7. The method for reducing noise of real-time speech at the edge according to claim 6, characterized in that: The first depthwise separable convolution layer includes a first depthwise convolution and a first pointwise convolution, and the second depthwise separable convolution layer includes a second depthwise convolution and a second pointwise convolution.

8. The method for reducing noise of real-time speech at the edge according to claim 6, characterized in that: The step of using a Wiener filter to perform adaptive iterative Wiener filtering on the speech frame and the non-speech frame to obtain a filtered frequency domain signal comprises: The average power spectrum of the silent segment is used as the initial noise power spectrum estimation, and the preset constant is used as the speech power spectrum; If the speech signal in the current frame is the non-speech frame, the noise power spectrum estimation is updated: in, The current frame The noise power spectrum estimation of frequency points, is the updated current frame The noise power spectrum estimation of frequency points, The current frame The power spectral density of the frequency point, is the smoothing coefficient; If the speech signal in the current frame is the speech frame, maintaining the noise power spectrum estimation unchanged; Calculate the gain of the Wiener filter based on the updated noise power spectrum estimate and speech power spectrum : The original frequency domain signal of the current frame is subjected to adaptive iterative Wiener filtering using the Wiener filter to obtain a filtered frequency domain signal; wherein the expression of the adaptive iterative Wiener filtering is: in, is the filtered frequency domain signal, is the original frequency domain signal.

Images