An adaptive noise reduction method and apparatus for a head-mounted computer

Abstract

The application provides an adaptive noise reduction method and device for a head-mounted computer, and relates to the technical field of voice noise reduction.The method comprises the following steps: acquiring a multi-channel signal in a rail transit operation environment, performing enhanced noise estimation on the multi-channel signal to obtain noise spectrum feature estimation, and collecting the multi-channel signal through a head-mounted computer; modeling background noise according to the noise spectrum feature estimation to obtain a noise source model; designing an adaptive spectrum filter based on the noise source model, filtering the multi-channel signal through the adaptive spectrum filter to obtain a preliminary noise reduction signal; and performing voice enhancement on the preliminary noise reduction signal based on a voice enhancement model to obtain a voice signal.The application solves the problem that adaptive noise reduction in a rail transit operation environment cannot simultaneously consider accuracy, robustness and real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech noise reduction, in particular to an adaptive noise reduction method and device for a head-mounted computer. BACKGROUND

[0002] In the rail transit operating environment, complex and diverse background noise (such as train operation noise, mechanical equipment sound, wind noise, etc.) brings great challenges to the extraction and enhancement of speech signals. These noises have the characteristics of wideband, non-stationary, complex spatial distribution, etc., and the spectral characteristics of the noise source will change over time in a dynamic environment, increasing the difficulty of signal processing.

[0003] Traditional noise reduction methods usually rely on fixed noise models, which are difficult to adapt to complex scenarios and can easily lead to distortion of speech signals or residual noise. In addition, speech enhancement needs to process both global frequency domain characteristics and local time-frequency characteristics, but existing methods often fail to take into account these characteristics, resulting in unstable noise reduction effects. The microphone array of a head-mounted computer provides a basis for multi-channel signal acquisition and processing, but its limited computing resources and energy consumption requirements place higher demands on the real-time and efficiency of the algorithm. How to design an adaptive noise reduction method that takes into account accuracy, robustness, and real-time performance in a complex environment has become a difficult problem to be solved in the field of speech processing in rail transit operating environments. SUMMARY

[0004] The purpose of the present application is to provide an adaptive noise reduction method and device for a head-mounted computer to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present application is as follows:

[0005] In a first aspect, the present application provides an adaptive noise reduction method for a head-mounted computer, comprising:

[0006] Obtaining a multi-channel signal in a rail transit operating environment, performing enhanced noise estimation on the multi-channel signal to obtain noise spectral feature estimation, and collecting the multi-channel signal through a head-mounted computer;

[0007] Modeling background noise based on the noise spectral feature estimation to obtain a noise source model;

[0008] Designing an adaptive spectral filter based on the noise source model, filtering the multi-channel signal through the adaptive spectral filter to obtain a preliminary noise reduction signal;

[0009] Performing speech enhancement on the preliminary noise reduction signal based on a speech enhancement model to obtain a speech signal.

[0010] In a second aspect, the present application also provides an adaptive noise reduction device for a head-mounted computer, comprising:

[0011] The acquisition module is used to acquire multi-channel signals in the rail transit operation environment, perform enhanced noise estimation on the multi-channel signals, and obtain noise spectrum characteristic estimation. The multi-channel signals are acquired by a head-mounted computer.

[0012] The module is used to model background noise based on the noise spectrum characteristics to obtain a noise source model;

[0013] The noise reduction module is used to design an adaptive spectrum filter based on the noise source model, and to filter the multi-channel signal through the adaptive spectrum filter to obtain a preliminary noise reduction signal;

[0014] An enhancement module is used to enhance the initial noise reduction signal based on a speech enhancement model to obtain a speech signal.

[0015] The beneficial effects of this invention are as follows: This invention utilizes a microphone array to acquire multi-channel signals, combines global frequency domain characteristics and time-frequency local characteristics to enhance noise estimation, and establishes a dynamic noise source model. Based on the noise source model, filter parameters are adaptively adjusted to effectively suppress the noise spectrum in real time, improving the quality of the initial noise-reduced signal. Furthermore, based on the initial noise-reduced signal, a speech enhancement model is used to extract the speech signal, making the speech clearer while maintaining its naturalness. Simultaneously, considering the hardware limitations of head-mounted computers, the algorithm structure is optimized to ensure real-time and efficient speech processing under limited resource conditions.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the adaptive noise reduction method for head-mounted computers described in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the adaptive noise reduction device for head-mounted computers described in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present 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 the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] Example 1:

[0023] This embodiment provides an adaptive noise reduction method for head-mounted computers.

[0024] See Figure 1 The figure shows that the method includes steps S100, S200, S300 and S400.

[0025] Step S100: Acquire multi-channel signals in the rail transit operation environment, perform enhanced noise estimation on the multi-channel signals, and obtain noise spectrum feature estimation. The multi-channel signals are acquired by a head-mounted computer.

[0026] In this embodiment, the noise in the rail transit operation environment includes low-frequency mechanical noise, wheel and track noise, and high-frequency ventilation and passenger noise. It is also necessary to consider that the noise in the rail transit environment is dynamically changing. In the rail transit operation environment, the noise exhibits characteristics of non-stationarity, broadband noise, multi-source noise, and overlap between background noise and speech signals.

[0027] Therefore, by enhancing noise estimation to separate the noise component from a mixed signal (i.e., speech and noise signals), the spectral characteristics of the noise can be obtained. Enhanced noise estimation allows us to understand the characteristics, intensity, and distribution of noise in the time and frequency domains, which helps in subsequently building dynamic noise source models.

[0028] Step S100 includes:

[0029] Step S101: Perform signal preprocessing on the high-frequency components of the multi-channel signal to obtain a preprocessed signal;

[0030] Step S102: Calculate the Mel frequency cepstral coefficients based on the global frequency domain characteristics and time-frequency local characteristics of the preprocessed signal;

[0031] Step S102 includes:

[0032] Step A100: Based on the microphone array of the head-mounted computer, beamforming is performed on the preprocessed signal to obtain an enhanced power spectrum;

[0033] Step A100 includes:

[0034] Step A101: Obtain the microphone array of the head-mounted computer;

[0035] Step A102: Perform a short-time Fourier transform on the preprocessed signal based on the microphone array to obtain the global time-frequency domain representation of different channels;

[0036] Step A103: Calculate the phase difference between different microphones based on the microphone array;

[0037] Step A104: Calculate the sound source direction estimate using the phase difference and minimum variance distortionless response method;

[0038] Step A105: Beamforming is performed on the preprocessed signal using the sound source direction estimation and the global time-frequency domain representation to obtain an enhanced signal, and the enhanced power spectrum of the enhanced signal is calculated.

[0039] In this embodiment, utilizing the spatial information of the microphone array can significantly improve the signal-to-noise ratio of the target speech and reduce noise interference from multiple directions during rail transit operations. Furthermore, estimating the sound source direction using the minimum variance distortionless response effectively eliminates noise from other directions. Beamforming can be combined with the global frequency domain characteristics of the short-time Fourier transform, integrating spatial and frequency domain information to improve overall noise reduction performance.

[0040] Step A200: Perform wavelet transform on the preprocessed signal to obtain the wavelet transform result;

[0041] Step A300: Calculate the frequency of the preprocessed signal and convert it to obtain a nonlinear Mel frequency;

[0042] Step A400: Calculate the Mel band energy based on the enhanced power spectrum, the wavelet transform result, and the Mel frequency;

[0043] Step A500: Perform a logarithmic energy transformation on the Mel band energy to obtain Mel logarithmic energy, wherein the Mel logarithmic energy includes enhanced Mel logarithmic energy and wavelet Mel logarithmic energy;

[0044] Step A600: Perform discrete cosine transform on the Mel logarithmic energy to obtain enhanced Mel frequency cepstral coefficients and wavelet Mel frequency cepstral coefficients;

[0045] Step A700: The enhanced Mel frequency cepstral coefficients and the wavelet Mel frequency cepstral coefficients are weighted and fused to obtain the Mel frequency cepstral coefficients of the preprocessed signal.

[0046] In step A700, the formula for calculating the Mel frequency cepstral coefficients is as follows:

[0047] MFCC k (t)=α·MFCC F,k (t)+β·MFCC W,k (t)

[0048]

[0049] In the formula, MFCC K (t) represents the k-th Mel-frequency cepstral coefficient, where α and β are both weighting coefficients, MFCC F,k (t) represents the k-th enhanced Mel frequency cepstral coefficient, MFCC W,k (t) represents the k-th wavelet Mel frequency cepstral coefficient, M represents the number of Mel filters, k represents the coefficient index parameter, m represents the Mel filter index parameter, and logE F,m (t) represents the enhanced Al logarithmic energy, logE F,m (t) represents the small Pomeranian logarithmic energy.

[0050] Step S103: Noise estimation is performed using the Mel frequency cepstral coefficients to obtain noise spectrum characteristic estimates.

[0051] In this embodiment, the calculation of the Mel frequency cepstral coefficients combines short-time Fourier transform and wavelet transform. Short-time Fourier transform can effectively analyze the frequency characteristics in rail transit operations and extract global frequency domain characteristics, such as low-frequency train noise and mid-to-high-frequency mechanical noise. Beamforming uses the phase and amplitude information of multi-channel signals through a microphone array to focus the target sound source signal while suppressing interference noise from other directions.

[0052] Wavelet transform performs multi-scale analysis of signals and is adaptable to non-stationary signals, enabling it to capture the details of low-frequency and high-frequency noise more accurately. It is particularly suitable for track noise and mechanical noise, which often exhibit significant local variations over time. Therefore, Mel-frequency cepstral coefficients based on beamforming and wavelet transform can effectively address complex noise scenarios in rail transit operations, improve the accuracy of noise spectral feature estimation, enhance the discriminability of low-frequency and mid-frequency speech signals, and help distinguish overlapping portions of speech and low-frequency noise.

[0053] Step S200: Based on the noise spectrum characteristics estimation, perform background noise modeling to obtain a noise source model;

[0054] In this embodiment, background noise modeling aims to accurately establish a spatiotemporal distribution model of noise to cope with the complex noise changes in the rail transit operation environment. The key to background noise modeling is to capture the dynamic characteristics of noise, including its frequency evolution, time-varying characteristics, and spatial distribution, so that the noise reduction algorithm can adaptively handle different noise sources, better adapt to the changing noise environment, and provide a more stable noise reduction effect.

[0055] Step S200 includes:

[0056] Step S201: Establish the probability distribution of noise spectral feature estimation based on the Gaussian mixture model, and select the number of Gaussian distributions through the Bayesian information criterion to obtain the preliminary noise source model;

[0057] In this embodiment, the Gaussian mixture model can effectively characterize the multimodal distribution characteristics of the noise spectrum and is suitable for complex noise scenarios in rail transit operation environments. At the same time, by selecting the number of Gaussian distributions through the Bayesian information criterion, the model complexity and fitting accuracy can be balanced, avoiding overfitting or underfitting.

[0058] In this embodiment, the expression for the probability distribution of noise spectrum feature estimation is:

[0059]

[0060] In the formula, p(x′) represents the probability distribution of the noise spectral feature estimation vector x′, A represents the number of Gaussian distributions, and π a μ represents the mixing weight of the a-th Gaussian component. a Let ∑ represent the mean vector of the a-th Gaussian component. a Let the covariance matrix of the a-th Gaussian component be denoted as . Let represent the probability density function of the a-th Gaussian component.

[0061] Step S202: Predict the noise spectrum characteristics at the current moment by using Kalman filtering to obtain preliminary prediction results, which include preliminary prediction state and preliminary prediction covariance;

[0062] In this embodiment, Kalman filtering can make real-time predictions of the noise spectrum characteristics at the current moment, and the preliminary prediction covariance can quantify the uncertainty of the prediction and enhance the reliability of the noise source model.

[0063] Step S203: Collect a preset number of particle swarms based on all the preliminary prediction results, where each particle in the particle swarm represents a preliminary prediction state;

[0064] Step S204: Update the particle weights based on the noise spectrum characteristics and probability distribution at the next time step;

[0065] Step S205: Resample the preliminary prediction results according to the particle weights to obtain a new particle swarm;

[0066] Step S206: Update the state and covariance of the preliminary prediction results based on the new particle swarm to obtain the updated noise source state;

[0067] Step S207: Adjust the mean and covariance of the preliminary noise source model by updating the noise source state to obtain the noise source model.

[0068] In this embodiment, the global optimization capability of the particle swarm can effectively handle the nonlinearity and non-Gaussianity in noise characteristics. Through weight updates and resampling, the particle swarm can dynamically adjust its distribution, thereby improving its ability to capture changes in the noise spectrum.

[0069] By combining Kalman filtering and particle swarm optimization, and making full use of historical prediction results and current observation information, the accuracy and robustness of noise source models can be improved. Furthermore, the models can be adjusted in real time according to the dynamic changes in the noise spectrum to adapt to different noise scenarios.

[0070] Step S300: Design an adaptive spectrum filter based on the noise source model, and filter the multi-channel signal using the adaptive spectrum filter to obtain a preliminary noise reduction signal;

[0071] Step S300 includes:

[0072] Step S301: Obtain the noise power spectral density using the noise source model;

[0073] Step S302: Select the minimum mean square error algorithm as the spectrum filter;

[0074] Step S303: Design the spectral weighting factor of the spectral filter based on the noise power spectral density;

[0075] Step S304: Design the weight update formula for the spectrum filter using the spectrum weighting factor to obtain the adaptive spectrum filter;

[0076] In step S304, the weight update formula is:

[0077] w(t+1)=w(t)+μ·α(f)·e(t)·x(t)

[0078]

[0079] In the formula, w(t+1) represents the weight at time t+1, w(t) represents the weight at time t, μ represents the step size factor, α(f) represents the spectral weighting factor with respect to frequency f, e(t) represents the error signal at time t, x(t) represents the input signal of the adaptive spectral filter, and P noise (f) represents the noise power spectral density at frequency f, and ∈ represents a constant parameter.

[0080] Step S305: The multi-channel signal is denoised using the adaptive spectrum filter to obtain a preliminary denoised signal.

[0081] In this embodiment, the weight update formula is designed by using a spectrum weighting factor, which enables the filter to have adaptive capabilities. The filter parameters can be dynamically adjusted according to the real-time changes in noise, thereby improving the flexibility and adaptability of noise reduction.

[0082] Step S400: Perform speech enhancement on the preliminary noise reduction signal based on the speech enhancement model to obtain a speech signal.

[0083] Step S400 includes:

[0084] Step S401: Construct a WaveNet model based on the preliminary noise reduction signal;

[0085] In this embodiment, the WaveNet model is an autoregressive generative model based on convolutional neural networks and causal convolution, capable of generating high-quality speech signals.

[0086] The formula for the WaveNet model is:

[0087] p(x t |x <t = WaveNet(x) <t )

[0088] In the formula, x t x represents the speech signal at the current moment. <t p(x) represents the speech signal prior to the current moment. t |x <t) represents a given x <t x is generated at time t The probability distribution at time t, WaveNet(·) represents the WaveNet model.

[0089] Step S402: Update the WaveNet model using an incremental learning method to obtain a speech enhancement model;

[0090] In this embodiment, the formula for updating the WaveNet model is:

[0091]

[0092] In the formula, θ t θ represents the WaveNet model parameters at the current time step. t+1 The WaveNet model parameters for the next time step, where η represents the learning rate. This represents the gradient of the loss function with respect to the WaveNet model parameters.

[0093] Step S403: Input the preliminary noise reduction signal into the speech enhancement model for speech enhancement to obtain a speech signal.

[0094] In this embodiment, to effectively address the computational resource requirements, real-time performance, and adaptability issues of the WaveNet model, an incremental learning method is incorporated. This allows the WaveNet model to continuously optimize in different environments and improve speech enhancement performance in complex noisy backgrounds. The incremental learning method enables the WaveNet model to adapt to new data over time without retraining the entire WaveNet model, thereby improving real-time processing capabilities and reducing computational resource consumption.

[0095] Example 2:

[0096] In this embodiment, the Wiener filter and the noise reduction method designed in this invention are selected, and experiments are conducted on the publicly available NOISEX-92 dataset, which contains speech signals under various noise environments. Signal-to-noise ratio, speech intelligibility, and speech quality are selected as evaluation metrics.

[0097] Table 1 shows a performance comparison between the Wiener filter and the noise reduction method of this invention, where SNR represents the signal-to-noise ratio, STOI represents speech intelligibility, and PESQ represents speech quality.

[0098] Table 1. Performance Comparison of Wiener Filter and Noise Reduction Method of the Present Invention

[0099] SNR (dB) STOI PESQ Original signal 5.0 0.45 1.7 Wiener filter 12.5 0.71 2.3 Noise reduction method of the invention 21.0 0.84 3.0

[0100] As shown in Table 1, the Wiener filter provides an improvement of approximately 7.5 dB, while the method of the present invention provides an improvement of 16 dB, demonstrating a significant noise reduction effect. At the same time, the method of the present invention also significantly improves STOI and PESQ, demonstrating better speech recovery effect and higher recovered speech quality.

[0101] In summary, the noise reduction method of the present invention has good speech noise reduction and enhancement effects in the complex noise background of rail transit operation environment, and can provide higher noise reduction accuracy and speech quality.

[0102] Example 3:

[0103] like Figure 2 As shown, this embodiment provides an adaptive noise reduction device for a head-mounted computer, the device comprising:

[0104] The acquisition module is used to acquire multi-channel signals in the rail transit operation environment, perform enhanced noise estimation on the multi-channel signals, and obtain noise spectrum characteristic estimation. The multi-channel signals are acquired by a head-mounted computer.

[0105] The module is used to model background noise based on the noise spectrum characteristics to obtain a noise source model;

[0106] The noise reduction module is used to design an adaptive spectrum filter based on the noise source model, and to filter the multi-channel signal through the adaptive spectrum filter to obtain a preliminary noise reduction signal;

[0107] An enhancement module is used to enhance the initial noise reduction signal based on a speech enhancement model to obtain a speech signal.

[0108] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can explore the technical scope disclosed in the present invention.

[0111] Any variations or substitutions that can be easily conceived within the scope of this invention should be included within the protection scope of this invention.

[0112] Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. An adaptive noise reduction method for head-mounted computers, characterized in that, include: Multi-channel signals from the rail transit operation environment are acquired, and enhanced noise estimation is performed on the multi-channel signals to obtain noise spectrum characteristic estimation. The multi-channel signals are acquired through a head-mounted computer. Background noise modeling is performed based on the noise spectrum feature estimation to obtain a noise source model, including: The probability distribution for noise spectral characteristic estimation is established based on the Gaussian mixture model, and the number of Gaussian distributions is selected through the Bayesian information criterion to obtain a preliminary noise source model. The expression for the probability distribution of noise spectrum feature estimation is: ； In the formula, A vector representing the noise spectrum feature estimation The probability distribution, Indicates the quantity of the Gaussian distribution. Indicates the first The mixing weights of Gaussian components, Indicates the first The mean vector of Gaussian components, Indicates the first The covariance matrix of Gaussian components, Indicates the first The probability density function of Gaussian components; The noise spectrum characteristics at the current moment are estimated and predicted by Kalman filtering to obtain preliminary prediction results, which include preliminary prediction state and preliminary prediction covariance. Based on all the preliminary prediction results, a preset number of particle swarms are collected, where each particle in the particle swarm represents a preliminary prediction state. The particle weights are updated based on the noise spectrum characteristics and probability distribution estimated at the next time step. The preliminary prediction results are resampled based on the particle weights to obtain a new particle swarm. Based on the new particle swarm, the state and covariance of the preliminary prediction results are updated to obtain the updated noise source state. The noise source model is obtained by adjusting the mean and covariance of the initial noise source model with the updated noise source state. An adaptive spectrum filter is designed based on the noise source model. The multi-channel signal is filtered by the adaptive spectrum filter to obtain a preliminary noise reduction signal. The initial noise-reduced signal is enhanced using a speech enhancement model to obtain a speech signal.

2. The adaptive noise reduction method for head-mounted computers according to claim 1, characterized in that... The multi-channel signal is subjected to enhanced noise estimation to obtain noise spectrum feature estimation, including: The high-frequency components of the multi-channel signal are preprocessed to obtain a preprocessed signal; The Mel frequency cepstral coefficients are calculated based on the global frequency domain characteristics and time-frequency local characteristics of the preprocessed signal; Noise spectral characteristics are estimated by using the Mel frequency cepstral coefficients.

3. The adaptive noise reduction method for head-mounted computers according to claim 2, characterized in that... The Mel frequency cepstral coefficients are calculated based on the global frequency domain characteristics and time-frequency local characteristics of the preprocessed signal, including: Based on the microphone array of the head-mounted computer, beamforming is performed on the preprocessed signal to obtain an enhanced power spectrum; Perform wavelet transform on the preprocessed signal to obtain the wavelet transform result; The frequency of the preprocessed signal is calculated and converted to obtain a nonlinear Mel frequency; Based on the enhanced power spectrum, the wavelet transform result, and the Mel frequency, the Mel band energy is calculated; The Mel band energy is subjected to a logarithmic energy transformation to obtain the Mel logarithmic energy, which includes enhanced Mel logarithmic energy and wavelet Mel logarithmic energy. Performing a discrete cosine transform on the Mel logarithmic energy yields enhanced Mel frequency cepstral coefficients and wavelet Mel frequency cepstral coefficients. The enhanced Mel frequency cepstral coefficients and the wavelet Mel frequency cepstral coefficients are weighted and fused to obtain the Mel frequency cepstral coefficients of the preprocessed signal.

4. The adaptive noise reduction method for head-mounted computers according to claim 3, characterized in that... Based on the microphone array of the head-mounted computer, beamforming is performed on the preprocessed signal to obtain an enhanced power spectrum, including: Obtain the microphone array of the head-mounted computer; Based on the microphone array, a short-time Fourier transform is performed on the preprocessed signal to obtain global time-frequency domain representations for different channels; Calculate the phase difference between different microphones based on the microphone array; The sound source direction estimate is calculated using the phase difference and minimum variance distortionless response method. Beamforming is performed on the preprocessed signal using the sound source direction estimation and the global time-frequency domain representation to obtain an enhanced signal, and the enhanced power spectrum of the enhanced signal is calculated.

5. The adaptive noise reduction method for head-mounted computers according to claim 3, characterized in that... The formula for calculating the Mel frequency cepstral coefficients is as follows: ； ； ； In the formula, Indicates the first Mel frequency cepstral coefficients, and All of these represent weighting coefficients. Indicates the first One enhanced Mel frequency cepstral coefficient, Indicates the first Small wavelet frequency cepstral coefficients Indicates the number of Mel filters. Indicates the coefficient index parameter. This represents the index parameter of the Mel filter. This indicates the enhancement of the Algebraic energy. This represents the small Pomeranian logarithmic energy.

6. The adaptive noise reduction method for head-mounted computers according to claim 1, characterized in that... An adaptive spectrum filter is designed based on the noise source model. This adaptive spectrum filter is used to filter multi-channel signals to obtain a preliminary noise-reduced signal, including: The noise power spectral density is obtained through the noise source model; The minimum mean square error algorithm was chosen as the spectrum filter. Design the spectral weighting factor of the spectral filter based on the noise power spectral density; An adaptive spectrum filter is obtained by designing the weight update formula for the spectrum filter using the aforementioned spectrum weighting factor. The multi-channel signal is denoised using the adaptive spectrum filter to obtain a preliminary denoised signal.

7. The adaptive noise reduction method for head-mounted computers according to claim 6, characterized in that, The weight update formula is as follows: ； ； In the formula, express Weight of time, express Weight of time, Indicates the step size factor. Indicates frequency Spectral weighting factor express Error signal at time, This represents the input signal of the adaptive spectrum filter. Indicates frequency The noise power spectral density, This represents a constant parameter.

8. The adaptive noise reduction method for head-mounted computers according to claim 1, characterized in that, Its features are, The initial noise-reduced signal is enhanced using a speech enhancement model to obtain a speech signal, including: A WaveNet model is constructed based on the initial noise-reduced signal; The WaveNet model is updated using an incremental learning method to obtain a speech enhancement model; The initial noise reduction signal is input into the speech enhancement model for speech enhancement to obtain a speech signal.

9. An adaptive noise reduction device for a head-mounted computer, characterized in that, include: The acquisition module is used to acquire multi-channel signals in the rail transit operation environment, perform enhanced noise estimation on the multi-channel signals, and obtain noise spectrum characteristic estimation. The multi-channel signals are acquired by a head-mounted computer. A construction module, used to model background noise based on the noise spectrum characteristics to obtain a noise source model, includes: The probability distribution for noise spectral characteristic estimation is established based on the Gaussian mixture model, and the number of Gaussian distributions is selected through the Bayesian information criterion to obtain a preliminary noise source model. The expression for the probability distribution of noise spectrum feature estimation is: ； In the formula, A vector representing the noise spectrum feature estimation The probability distribution, Indicates the quantity of the Gaussian distribution. Indicates the first The mixing weights of Gaussian components, Indicates the first The mean vector of Gaussian components, Indicates the first The covariance matrix of Gaussian components, Indicates the first The probability density function of Gaussian components; The noise spectrum characteristics at the current moment are estimated and predicted by Kalman filtering to obtain preliminary prediction results, which include preliminary prediction state and preliminary prediction covariance. Based on all the preliminary prediction results, a preset number of particle swarms are collected, where each particle in the particle swarm represents a preliminary prediction state. The particle weights are updated based on the noise spectrum characteristics and probability distribution estimated at the next time step. The preliminary prediction results are resampled based on the particle weights to obtain a new particle swarm. Based on the new particle swarm, the state and covariance of the preliminary prediction results are updated to obtain the updated noise source state. The noise source model is obtained by adjusting the mean and covariance of the initial noise source model with the updated noise source state. The noise reduction module is used to design an adaptive spectrum filter based on the noise source model, and to filter the multi-channel signal through the adaptive spectrum filter to obtain a preliminary noise reduction signal; An enhancement module is used to enhance the initial noise reduction signal based on a speech enhancement model to obtain a speech signal.

Images