Noise reduction method and system for dual-microphone neural network noise reduction headphones

Abstract

The present invention relates to a noise reduction method and system for dual-microphone neural network noise reduction headphones, comprising the following steps: collecting sound signals through two microphones of headphones to obtain original sound data; performing preliminary separation on the original sound data to obtain a first signal and a second signal; respectively calculating the energy difference and cross-power spectrum of the first signal and the second signal, calculating the transfer function between the first signal and the second signal according to the energy difference and the cross-power spectrum, calculating the gain function in the frequency domain according to the energy difference, the cross-power spectrum and the transfer function, multiplying the first signal by the gain function to obtain an enhanced desired audio signal; processing the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal. It can ensure that the main speaker's voice is not distorted, and all surrounding noise and interference are suppressed to the greatest extent.

Description

Technical Field

[0001] The present invention relates to the technical field of noise reduction headphones, and in particular to a noise reduction method and system for dual-microphone neural network noise reduction headphones. Background Art

[0002] In today's open space office scene with multiple people, it is difficult to ensure a quiet environment. Especially in meetings, the two parties on the call often cannot accurately obtain information due to the surrounding environmental noise (air conditioning, conversation, etc.), making communication difficult and inefficient. Summary of the invention

[0003] To this end, the technical problem to be solved by the present invention is to overcome the technical problems in the prior art that accurate information cannot be obtained due to the influence of ambient sound, communication is difficult and inefficient.

[0004] In order to solve the above technical problems, the present invention provides a noise reduction method for dual-microphone neural network noise reduction headphones, comprising the following steps:

[0005] S1, collecting sound signals through two microphones of the headset to obtain original sound data;

[0006] S2. Preliminarily separate the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of the main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal;

[0007] S3, respectively calculating the energy difference and the cross power spectrum of the first signal and the second signal, calculating the transfer function between the first signal and the second signal according to the energy difference and the cross power spectrum, obtaining a gain function in the frequency domain according to the energy difference, the cross power spectrum and the transfer function, and multiplying the first signal by the gain function to obtain an enhanced desired audio signal;

[0008] S4. Processing the enhanced desired audio signal by a voice activity detection algorithm to obtain a clean main speaker's voice signal.

[0009] Preferably, in S2, the original sound data is preliminarily separated to obtain a first signal and a second signal, including:

[0010] Performing fast Fourier transform on the sound signals collected by the two microphones respectively, separating the sound signals into different sub-bands, and obtaining a first frequency domain signal and a second frequency domain signal;

[0011] Construct an initial separation matrix B(0, f) on each subband;

[0012] Calculate the natural gradient ΔB(k,f) of each subband in the frequency domain;

[0013] According to the natural gradient of each subband, the separation matrix is ​​updated, and the updated separation matrix is:

[0014] B(k,f)=B(k-1,f)+μΔB(k,f),

[0015] Among them, k is the number of frames, f represents the center frequency of each sub-band, and μ is the step size, which generally ranges from 0.01 to 0.1. Filter 11 is 1, filter 12 is 0, filter 21 is 1, filter 22 is 0;

[0016] According to the updated separation matrix, the frequency domain output of blind source separation is calculated:

[0017]

[0018] Preferably, the components between S1 and S2 include:

[0019] Pre-emphasize the original sound data to obtain a pre-emphasized original signal.

[0020] Preferably, the step between S1 and S2 also includes:

[0021] Using a Hamming window to window the original sound data to obtain a windowed original signal;

[0022] Among them, the Hamming window function is as follows:

[0023] W(n)=0.54-0.46×cos(2×π×n / (N-1)), 0≤n≤N-1, where W is the window function, n is the number of sampling points of a frame signal, N is the frame length, cos is the cosine function, and π is the circumference of a circle.

[0024] Preferably, calculating the cross power spectrum of the first signal and the second signal in S3 includes:

[0025] Acquire the power of the first signal and the power of the second signal;

[0026] The conjugate of the power of the second signal is multiplied by the power of the first signal, and the absolute value is taken to obtain a cross-power spectrum of the first signal and the second signal.

[0027] Preferably, in S4, the voice activity detection algorithm is a voice activity detection algorithm based on a neural network.

[0028] Preferably, the neural network in S4 includes a convolutional layer, a PRelu activation layer, a maximum pooling layer, a normalization layer, an LSTM layer, a DNN fully connected layer and a sigmoid function layer arranged in sequence.

[0029] The present invention discloses a noise reduction system for dual-microphone neural network noise reduction headphones, comprising:

[0030] A signal acquisition module, wherein the signal acquisition module collects sound signals through two microphones of the earphone to obtain original sound data;

[0031] A blind source signal separation module, wherein the blind source signal separation module performs preliminary separation on the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of a main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal;

[0032] an expected audio acquisition module, the expected audio acquisition module being used to respectively calculate an energy difference and a cross-power spectrum of the first signal and the second signal, calculate a transfer function between the first signal and the second signal according to the energy difference and the cross-power spectrum, obtain a gain function in the frequency domain according to the energy difference, the cross-power spectrum and the transfer function, and multiply the first signal by the gain function to obtain an enhanced expected audio signal;

[0033] A detection module processes the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal.

[0034] Preferably, it also includes a data processing module, which pre-emphasizes the original sound data to obtain a pre-emphasized signal, uses a Hamming window to window the pre-emphasized signal to obtain the windowed original signal, and the blind source signal separation module processes the windowed original signal.

[0035] The above technical solution of the present invention has the following advantages compared with the prior art:

[0036] The present invention 1. The present invention can ensure that the main speaker's voice is not distorted while maximally suppressing all surrounding noise and interference, that is, retaining the complete voice information of the current speaker while completely shielding other surrounding sounds (air conditioning noise, conversation sounds of other speakers).

[0037] 2. The present invention is a dual-microphone system, which can ensure that the main microphone signal is not distorted, facilitate communication, improve work efficiency, and enable both parties of the call to clearly listen to the pure voice signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the noise reduction method of the dual-microphone neural network noise reduction headset in the present invention;

[0039] Figure 2 It is a structural diagram of a blind source signal separation module;

[0040] Figure 3 It is a structural diagram of the expected audio acquisition module;

[0041] Figure 4 Schematic diagram of the neural network structure of the detection module. DETAILED DESCRIPTION

[0042] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0043] Reference Figure 1 As shown, the present invention discloses a noise reduction method for dual-microphone neural network noise reduction headphones, comprising the following steps:

[0044] Step 1: Collect sound signals through the two microphones of the headset to obtain original sound data.

[0045] The method includes: pre-emphasizing the original sound data to obtain the pre-emphasized original signal; windowing the original sound data using a Hamming window to obtain the windowed original signal; wherein the Hamming window function is as follows: W(n)=0.54-0.46×cos(2×π×n / (N-1)), 0≤n≤N-1, wherein W is the window function, n is the number of sampling points of a frame signal, N is the frame length, cos is the cosine function, and π is the circumference of a circle.

[0046] Step 2: Preliminarily separate the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of the main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal.

[0047] The initial separation of the original sound data to obtain the first signal and the second signal includes:

[0048] Performing fast Fourier transform on the sound signals collected by the two microphones respectively, separating the sound signals into different sub-bands, and obtaining a first frequency domain signal and a second frequency domain signal;

[0049] Construct an initial separation matrix B(0, f) on each subband;

[0050] Calculate the natural gradient ΔB(k,f) of each subband in the frequency domain;

[0051] According to the natural gradient of each subband, the separation matrix is ​​updated, and the updated separation matrix is:

[0052] B(k,f)=B(k-1,f)+μΔB(k,f),

[0053] Among them, k is the number of frames, f represents the center frequency of each sub-band, and μ is the step size, which generally ranges from 0.01 to 0.1. Filter 11 is 1, filter 12 is 0, filter 21 is 1, filter 22 is 0;

[0054] According to the updated separation matrix, the frequency domain output of blind source separation is calculated:

[0055]

[0056] Step three, respectively calculate the energy difference and cross-power spectrum of the first signal and the second signal, calculate the transfer function between the first signal and the second signal according to the energy difference and the cross-power spectrum, obtain a gain function in the frequency domain according to the energy difference, the cross-power spectrum and the transfer function, and multiply the first signal by the gain function to obtain an enhanced desired audio signal.

[0057] Calculating the cross power spectrum of the first signal and the second signal in step three includes: acquiring the power of the first signal and the power of the second signal; multiplying the conjugate of the power of the second signal by the power of the first signal and taking the absolute value to obtain the cross power spectrum of the first signal and the second signal.

[0058] Step 4: Process the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal.

[0059] In step 4, the voice activity detection algorithm is a voice activity detection algorithm based on a neural network. The neural network includes a convolution layer, a PRelu activation layer, a maximum pooling layer, a normalization layer, an LSTM layer, a DNN fully connected layer and a sigmoid function layer arranged in sequence.

[0060] The invention discloses a noise reduction system for a dual-microphone neural network noise reduction headset, comprising a signal acquisition module, a blind source signal separation module, a desired audio acquisition module and a detection module.

[0061] The signal acquisition module collects sound signals through two microphones of the earphone to obtain original sound data;

[0062] The blind source signal separation module performs preliminary separation on the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of the main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal;

[0063] The desired audio acquisition module is used to calculate the energy difference and the cross power spectrum of the first signal and the second signal respectively, calculate the transfer function between the first signal and the second signal according to the energy difference and the cross power spectrum, obtain a gain function in the frequency domain according to the energy difference, the cross power spectrum and the transfer function, and multiply the first signal by the gain function to obtain an enhanced desired audio signal;

[0064] The detection module processes the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal.

[0065] The present invention also includes a data processing module, which pre-emphasizes the original sound data to obtain a pre-emphasized signal, uses a Hamming window to window the pre-emphasized signal to obtain a windowed original signal, and the blind source signal separation module processes the windowed original signal.

[0066] The technical solution of the present invention is further described and explained below in conjunction with specific embodiments.

[0067] Figure 2 As shown, it is a schematic diagram of a blind source signal separation module, which refers to separating the original sound data collected by the two microphones of the headset to achieve the purpose of separating the ambient sound and the main speaker's voice; wherein, the main process of blind source separation includes: pre-emphasis, framing, windowing, short-time fast Fourier transform (STFT) of the speech signal, calculation of the natural gradient of different STFT subbands, minimum distortion processing, short-time inverse fast Fourier transform (ISTFT), overlap-add, etc. As the first module of the present invention, after the original speech signal passes through the module, two separated signals in the frequency domain will be obtained, that is, BSS output 1 is the frequency domain signal of the main speaker's voice plus a small amount of interference and noise, and BSS output 2 is the interference signal and noise signal plus a small amount of the main speaker's signal. Due to the relationship between the presence of multiple sound sources and reverberation in the space, the blind source separation module cannot completely obtain the pure desired signal and interference signal, that is, some interference signals will remain in the desired signal. The main function of the blind source separation module is to provide preprocessing for the dual-microphone system and the next PLD module, preliminarily distinguish the main speaker from the interference, and accelerate the convergence speed and calculation accuracy of the subsequent modules. The specific process is as follows:

[0068] 1. Pre-emphasis: The pre-emphasis factor is usually 0.97, that is, each sampling point minus 97% of the previous sampling point. After pre-emphasis processing, the pre-emphasized signal is obtained.

[0069] 2. Windowing: To prevent frequency leakage, use the Hamming window for windowing. The Hamming window function is as follows:

[0070] W(n)=0.54-0.46×cos(2×π×n / (N-1)),0≤n≤N-1

[0071] The windowed signal is obtained by passing the window function. Where W is the window function, n is the number of sampling points of a frame signal, N is the frame length, cos is the cosine function, and π is the circumference of a circle.

[0072] 3. STFT: Perform short-time fast Fourier transform to separate the signal into different sub-bands. After fast Fourier transform, the window signal in the time domain is transformed into a signal in the frequency domain. Frequency domain signal 1 and frequency domain signal 2 are obtained. At the same time, the convolution operation in the time domain is transformed into a product operation in the frequency domain.

[0073] 4. Construct the initial separation matrix on each subband Where f is the center frequency of each subband, filter 11 is 1, filter 12 is 0, filter 21 is 1, and filter 22 is 0.

[0074] 5. Calculate the natural gradient and calculate the gradient ΔB(k,f) for each subband in the frequency domain.

[0075] 6. Update the separation matrix: According to the natural gradient of each subband, the separation matrix is ​​gradually updated from the initial separation matrix as follows:

[0076] B(k,f)=B(k-1,f)+μΔB(k,f)

[0077] The final separation matrix in the frequency domain is obtained. Where k is the number of frames, f represents the center frequency of each subband, and μ is the step size, which generally ranges from 0.01 to 0.1. B is the separation matrix, expressed as

[0078] 7. Calculate output: According to the updated separation matrix, calculate the frequency domain output of the current blind source separation as follows:

[0079]

[0080] Figure 3This is a schematic diagram of the expected audio acquisition module, which refers to the energy difference between the noisy speech signals picked up by the two microphones in a dual-microphone system. The input of this module is the output of the previous blind source separation module. After the signal passes through the blind source separation module, two preliminary separated signals are obtained. After the preliminary separated signal passes through this module, a clean signal with only the main speaker will be obtained, which is the PLD output of this module. In a scattered noise field and far-field noise environment, it can usually be considered that the speech signals reaching the two microphones have a large energy level difference, while the noise signals reaching the two microphones have only a small energy difference. The PLD algorithm uses this feature to enhance the expected speech signal. In order to keep the expected speech signal undistorted, a Wiener filter is constructed using the dual-microphone energy difference and the estimated transfer function to handle stationary and non-stationary processes. The specific process is as follows,

[0081] 1. Calculate the energy and cross-power spectrum of the two frequency domain blind source separation outputs respectively.

[0082] 2. Calculate the energy difference between the two signals based on the calculated energy.

[0083] 3. Calculate the transfer function between the two signals through frequency domain signals, energy difference and cross power spectrum.

[0084] 4. Calculate the final frequency domain gain function based on the variables obtained in the above three steps.

[0085] 5. Multiply the final frequency domain gain function by the bss output 1 to obtain the desired audio output.

[0086] The detection module is an important part of the voice front-end algorithm. Its purpose is to detect the voice from the audio signal obtained by the microphone so that the subsequent algorithm can process it. In real-time conference scenarios, the accuracy of the VAD algorithm has a great impact on the subsequent algorithm and the final sound quality. The traditional VAD method is mainly based on the characteristics of speech for modeling. It has high requirements for the external environment and speech signal-to-noise ratio, and cannot handle some transient noises such as knocking and keyboard sounds. In recent years, the VAD method based on neural networks has become increasingly popular. The powerful data fitting ability of neural networks is used to achieve voice detection in complex scenarios, and the effect is usually better than traditional algorithms.

[0087] Figure 4The neural network structure diagram of the present invention is shown, and the speech detection model is obtained by training the neural network. As shown in the figure, the frequency domain output of the PLD module passes through this module. First, the 40-dimensional features extracted by the feature extraction step are sent to the first layer of the model, the convolution layer. The convolution layer consists of 16 convolution kernels, each of which has a size of 1×8. The convolution kernel is convolved on the time-frequency axis. This step is to learn the association information between frequency sub-bands, and then the PRelu activation function is used for calculation, and then the maximum pooling layer is connected, and the pooling size is 1×3. Then the pooled output is sent to the normalization layer, and the normalization layer normalizes each feature map, which can effectively reduce the occurrence of misjudgment caused by changes in speech amplitude. Then the output is sent to the LSTM layer, and the LSTM can effectively learn the association information between frames, which has a great effect on improving the accuracy of speech detection. Finally, it is sent to the DNN fully connected layer for classification, and the final frame prediction result is output through the sigmoid function. This module is the last module of the present invention, which is responsible for detecting the knocking sound, keyboard sound, etc. in the clean speech signal. The final output of this module is the clean audio signal of the main speaker.

[0088] According to tests, the present invention can eliminate all environmental noise beyond 20cm around the speaker, and the signal-to-interference ratio reaches -75dB.

[0089] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0090] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0091] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0093] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A noise reduction method for dual-microphone neural network noise reduction headphones, characterized in that: The following steps are involved: S1, collecting sound signals through two microphones of the headset to obtain original sound data; S2. Preliminarily separate the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of the main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal; S3, respectively calculating the energy difference and the cross power spectrum of the first signal and the second signal, calculating the transfer function between the first signal and the second signal according to the energy difference and the cross power spectrum, obtaining a gain function in the frequency domain according to the energy difference, the cross power spectrum and the transfer function, and multiplying the first signal by the gain function to obtain an enhanced desired audio signal; S4, processing the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal; The step S2, in which the original sound data is preliminarily separated to obtain a first signal and a second signal, includes: Performing fast Fourier transform on the sound signals collected by the two microphones respectively, separating the sound signals into different sub-bands, and obtaining a first frequency domain signal and a second frequency domain signal; Construct an initial separation matrix B(0, f) on each subband; Calculate the natural gradient ΔB(k, f) of each subband in the frequency domain; According to the natural gradient of each subband, the separation matrix is ​​updated, and the updated separation matrix is: B(k,f)=B(k-1,f)+μΔB(k,f), Where k is the number of frames, f represents the center frequency of each subband, and μ is the step size, ranging from 0.01 to 0.1; According to the updated separation matrix, the frequency domain output of blind source separation is calculated.

2. The noise reduction method of the dual-microphone neural network noise reduction headset according to claim 1, characterized in that: The S1 and S2 include: Pre-emphasize the original sound data to obtain a pre-emphasized original signal.

3. The noise reduction method of the dual-microphone neural network noise reduction headset according to claim 1, characterized in that: The S1 and S2 also include: Using a Hamming window to window the original sound data to obtain a windowed original signal; Among them, the Hamming window function is as follows: W(n)=0.54-0.46×cos(2×π×n / (N-1)), 0≤n≤N-1, where W is the window function, n is the number of sampling points of a frame signal, N is the frame length, cos is the cosine function, and π is the circumference of a circle.

4. The noise reduction method of the dual-microphone neural network noise reduction headset according to claim 1, characterized in that: The step of calculating the cross power spectrum of the first signal and the second signal in S3 includes: Acquire the power of the first signal and the power of the second signal; The conjugate of the power of the second signal is multiplied by the power of the first signal, and the absolute value is taken to obtain the cross-power spectrum of the first signal and the second signal.

5. The noise reduction method of the dual-microphone neural network noise reduction headset according to claim 1, characterized in that: In S4, the voice activity detection algorithm is a voice activity detection algorithm based on a neural network.

6. The noise reduction method of the dual-microphone neural network noise reduction headset according to claim 5, characterized in that: The neural network in S4 includes a convolutional layer, a PRelu activation layer, a maximum pooling layer, a normalization layer, an LSTM layer, a DNN fully connected layer and a sigmoid function layer which are arranged in sequence.

7. A noise reduction system for dual-microphone neural network noise reduction headphones, characterized in that: include: A signal acquisition module, wherein the signal acquisition module collects sound signals through two microphones of the earphone to obtain original sound data; A blind source signal separation module, wherein the blind source signal separation module performs preliminary separation on the original sound data to obtain a first signal and a second signal, wherein both the first signal and the second signal include an interfering frequency domain signal and a frequency domain signal of a main speaker's voice; wherein the proportion of the frequency domain signal of the main speaker's voice in the first signal is greater than the interfering frequency domain signal, and the proportion of the frequency domain signal of the main speaker's voice in the second signal is less than the interfering frequency domain signal; an expected audio acquisition module, the expected audio acquisition module being used to respectively calculate an energy difference and a cross-power spectrum of the first signal and the second signal, calculate a transfer function between the first signal and the second signal according to the energy difference and the cross-power spectrum, obtain a gain function in the frequency domain according to the energy difference, the cross-power spectrum and the transfer function, and multiply the first signal by the gain function to obtain an enhanced expected audio signal; A detection module, wherein the detection module processes the enhanced desired audio signal through a voice activity detection algorithm to obtain a clean main speaker's voice signal; The blind source signal separation module performs preliminary separation on the original sound data to obtain the first signal and the second signal, including: Performing fast Fourier transform on the sound signals collected by the two microphones respectively, separating the sound signals into different sub-bands, and obtaining a first frequency domain signal and a second frequency domain signal; Construct an initial separation matrix B(0, f) on each subband; Calculate the natural gradient ΔB(k, f) of each subband in the frequency domain; According to the natural gradient of each subband, the separation matrix is ​​updated, and the updated separation matrix is: B(k,f)=B(k-1,f)+μΔB(k,f), Where k is the number of frames, f represents the center frequency of each subband, and μ is the step size, ranging from 0.01 to 0.1; According to the updated separation matrix, the frequency domain output of blind source separation is calculated.

8. The noise reduction system of the dual-microphone neural network noise reduction headset according to claim 7, characterized in that: It also includes a data processing module, which pre-emphasizes the original sound data to obtain a pre-emphasized signal, uses a Hamming window to window the pre-emphasized signal to obtain the windowed original signal, and the blind source signal separation module processes the windowed original signal.

Images