Noise suppression method for spaceflight voice interaction equipment

Abstract

The invention discloses a noise suppression method for spaceflight voice interaction equipment, and the method comprises the steps: S1, processing an original voice signal through a spatial filter, and obtaining a first voice signal with spatial directivity; s2, performing post-filtering processing on the first voice signal, and suppressing low-frequency noise of the first voice signal to obtain a second voice signal; s3, inverse filtering processing is carried out on the second voice signal, small-space transmission response of the second voice signal is suppressed, and a third voice signal is obtained; and S4, suppressing nonlinear noise in the third voice signal by using a nonlinear mapping network to obtain a fourth voice signal suitable for spaceflight voice communication. According to the invention, for a complex noise environment in a spacecraft cabin, effective multi-channel voice enhancement and noise suppression effects can be achieved, the lightweight level of spaceflight voice interaction equipment is improved, and the complexity of a voice interaction system is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of manned spaceflight speech technology, and in particular relates to a noise suppression method for spaceflight speech interaction equipment. Background Art

[0002] In modern space missions, voice interaction devices play an increasingly important role. The devices can provide astronauts with various voice interaction services such as voice communication and voice recognition, thus meeting the increasingly complex, sophisticated and intelligent space mission requirements.

[0003] Due to the limitations of equipment and materials in the spacecraft cabin, the working environment of astronauts today is still noisy. It mainly comes from various equipment and instruments inside the spacecraft, such as the power system, ventilation system, water circulation system, etc. These devices provide necessary life support for astronauts or maintain daily working conditions, and generally need to operate continuously. Therefore, astronauts have to face the influence of complex noise environment when working.

[0004] Noise has a great impact on the voice interaction system used in aerospace. Both the quality of voice communication and the recognition rate of voice recognition are closely related to the environmental noise. Considering the special requirements of high reliability and high robustness in aerospace applications, in order to effectively deal with the constraints of the complex noise environment in the spacecraft cabin on the voice interaction performance, the voice processor in the spacecraft must have noise suppression capabilities and integrate effective noise suppression technology for the spacecraft cabin environment.

[0005] Traditional single-channel noise suppression technology can be implemented quickly, and to a certain extent, it can achieve the purpose of noise reduction and improve speech quality. However, this processing often causes distortion of speech signals, reduces the intelligibility of speech, and cannot suppress noise well because it cannot collect spatial information of the signal. The microphone array can perform spatial filtering on noise signals, and can form beams in specific directions in complex speech environments to obtain speech signals emitted by specific sound sources, thereby achieving a more ideal noise suppression effect; at the same time, for specific enclosed spaces and transmission media in the spacecraft cabin, the small space transmission response is collected for inverse filtering, and a nonlinear mapping network is trained for noise suppression. However, considering the severely limited computing resources of aerospace voice processors, there is an urgent need to propose a low-computational-overhead, high-performance multi-channel noise suppression method for the spacecraft cabin environment. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned defects and provide a noise suppression method for aerospace voice interaction equipment, which solves the technical problem of the difficulty of noise suppression in complex noise environments. In the face of the complex noise environment in the spacecraft cabin, the present invention can play an effective role in multi-channel voice enhancement and noise suppression, improve the lightweight level of aerospace voice interaction equipment, and reduce the complexity of the voice interaction system.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0008] A noise suppression method for aerospace voice interaction equipment, comprising:

[0009] S1 processes the original speech signal using a spatial filter to obtain a first speech signal with spatial directivity;

[0010] S2 performs post-filtering processing on the first speech signal to suppress low-frequency noise of the first speech signal to obtain a second speech signal;

[0011] S3 performs inverse filtering on the second voice signal to suppress a small space transmission response of the second voice signal to obtain a third voice signal;

[0012] S4 utilizes a nonlinear mapping network to suppress nonlinear noise in the third voice signal to obtain a fourth voice signal suitable for aerospace voice communication.

[0013] Furthermore, the method of step S1 includes:

[0014] S1.1 presets the direction according to the previously collected sound source, and determines the optimized spatial filter weight parameters through the pre-trained spatial filter model;

[0015] S1.2 estimates the delay of the acquired original multi-channel speech signal according to a preset direction to obtain a multi-channel delay signal;

[0016] S1.3 performs delay compensation on multi-channel delay signals to align the multi-channel delay signals;

[0017] S1.4 uses the optimized spatial filtering weight parameters to perform weighted summation on the aligned multi-channel time-delayed signals to obtain a single-channel signal, that is, a first speech signal with spatial directivity.

[0018] Further, the method of step S2 includes:

[0019] Acquire noisy speech corpus in advance, and acquire denoised corpus under ideal conditions; the noisy speech corpus acquired in advance is preferably noisy speech corpus in a closed small space sound field of a specific structure of a spacecraft;

[0020] Obtaining corresponding ideal post-filter parameters according to the noisy speech corpus and the ideal denoised corpus;

[0021] The first speech signal is processed using the post-filter parameters under ideal conditions to obtain a second speech signal.

[0022] Furthermore, in step S3, the filter used when inverse filtering is performed on the second speech signal includes a plurality of short filters.

[0023] Furthermore, in step S4, the nonlinear mapping network is trained using the priori noise information obtained in advance, and the nonlinear noise in the third speech signal is suppressed using the trained nonlinear mapping network.

[0024] Further, the method of step S2 includes:

[0025] S2.1 constructs a Wiener filter parameter correction table and a table lookup classifier;

[0026] The Wiener filter parameter correction table is used to describe the correspondence between the noise corpus and the correction rules;

[0027] The table lookup classifier is used to describe the correspondence between the statistical values ​​of the acoustic representation of a number of consecutive frames without speech and the correction rules;

[0028] S2.2 calculating the posterior signal-to-noise ratio and the prior signal-to-noise ratio using the speaking frame in the first speech signal to obtain the posterior signal-to-noise ratio and the prior signal-to-noise ratio of the first speech signal in the frame, and obtaining the transfer function of the first speech signal in the frame according to the posterior signal-to-noise ratio and the prior signal-to-noise ratio, that is, the filter parameters of the first speech signal in the frame;

[0029] S2.3 calculating the statistical value of the acoustic representation of a plurality of consecutive speechless segments in the first speech signal, and obtaining the corresponding correction rule according to the table lookup classifier obtained in step S2.1;

[0030] S2.4 Based on the Wiener filter parameter correction table obtained in step S2.1, the filter parameters of the first speech signal obtained in step S2.2 are corrected using the correction rule obtained in step S2.3 to obtain corrected filter parameters;

[0031] S2.5 performs post-filtering processing on the first speech signal using the modified filter parameters obtained in step S2.4 to obtain a second speech signal.

[0032] Furthermore, in step S2.1, the statistical values ​​of the acoustic representation of a plurality of consecutive frames without speech include the mean and variance of the MFCC cepstrum features of the silent frames.

[0033] Furthermore, in step S3, a frequency-dependent linear prediction filter is first obtained through a series of historical frame information, and then the enhanced speech signal is obtained in the subband domain through the difference between the original response signal and the filtered signal.

[0034] Furthermore, in step S3, Kronecker integral decomposition is introduced to decompose the multi-channel linear prediction filter into three sub-filters.

[0035] Furthermore, in step S1.2, the method for determining the preset direction includes:

[0036] The direction of the maximum output power pointed by the multi-channel speech signal beam collected by the microphone array is used as the preset direction.

[0037] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0038] The present invention provides a noise suppression method for a space-use voice interaction device. Facing the complex noise environment in a spacecraft cabin, the present invention can play an effective role in multi-channel voice enhancement and noise suppression, improve the quality of voice communication, enhance the performance of a voice interaction system, and greatly enhance the practicality of the voice interaction system. The present invention also integrates the noise suppression method into an algorithm processing control submodule of a voice interaction device, thereby improving the lightweight level of a space-use voice interaction device and reducing the complexity of the voice interaction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an application structure diagram of a noise suppression method for aerospace voice interaction equipment of the present invention;

[0040] Figure 2 The present invention is an algorithm flow chart of a noise suppression method for aerospace voice interaction equipment;

[0041] Figure 3 The present invention is a flow chart of a delay-sum beamforming method based on spatial filtering weight parameters of a priori sound source distribution;

[0042] Figure 4 It is a flow chart of a method for adaptively correcting parameters of a post-Wiener filter based on a priori parameter correction table of the present invention;

[0043] Figure 5 It is a diagram of the parameter correction process of the post-Wiener filter based on the a priori parameter correction table of the present invention;

[0044] Figure 6 It is a flow chart of the weighted prediction error method based on a priori transmission response of the present invention;

[0045] Figure 7 This is a flow chart of the nonlinear mapping network method based on pre-training of the present invention. DETAILED DESCRIPTION

[0046] The following detailed description of the present invention will make the features and advantages of the present invention more clear and explicit.

[0047] The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise noted.

[0048] The present invention provides a low-computational-overhead, high-performance multi-channel noise suppression method which is mainly used in aerospace voice interaction equipment. During the microphone array voice acquisition process, the method can effectively reduce the influence of complex environmental noise in the spacecraft cabin on the voice interaction system while occupying only a small amount of computing resources, thereby achieving the purpose of improving voice quality and enhancing voice processing performance.

[0049] In view of the special acoustic environment in the spacecraft cabin (specific enclosed space and transmission medium), voice signals are collected in a real simulation environment to obtain characteristic information of the actual application scenario, including spatial information, noise distribution and transmission response. Based on the above prior information, the original voice signal is processed by spatial filtering, post-filtering, inverse filtering, nonlinear mapping, etc., so as to effectively suppress the noise caused by the special environment in the spacecraft cabin. The specific method is as follows:

[0050] (1) The original speech signal is processed by a spatial filter to obtain a first speech signal, making it spatially directional. Since the noise suppression method examples of the present invention all work in a fixed spacecraft, the direction of the signal source can be preset by a large number of sound sources (including speech signals and noise signals) collected and analyzed in advance, so as to determine the optimization parameters through a pre-trained spatial filtering model. After the original speech signal is processed by a spatial filter with optimized weight parameters, a first speech signal is obtained, which can suppress interference signals in non-target directions and enhance speech signals in the target direction;

[0051] (2) Processing the first speech signal by post-filtering to obtain the second speech signal and suppress its low-frequency noise. The above-mentioned first speech signal is a signal obtained by spatial filtering, but it is highly dependent on the direction of the signal source, and the suppression effect on low-frequency noise is not ideal, and a large number of microphones are required to achieve a better noise suppression effect. In response to the above-mentioned problem that restricts the noise suppression effect of spatial filtering, the present invention improves it through a post-filtering method. Similar to the above method, since each noise reduction algorithm instance works in a fixed spacecraft, a large amount of noisy speech data of the scene and the noise-reduced data under ideal conditions obtained by complex technical means can be obtained in advance, and then the corresponding ideal post-filter parameters can be estimated. After the first speech signal is processed by the parameter-corrected post-filter, the second speech signal is obtained, which can further enhance the low-frequency noise suppression effect;

[0052] (3) The second voice signal is processed by inverse filtering to obtain a third voice signal, and its small space transmission response is suppressed. Since the spacecraft cabin has a specific enclosed space and a transmission medium, the transmission response of the voice signal in its small space is quite different from that in other common scenarios. In view of the above situation, the present invention collects and analyzes the transmission response of the actual application scenario and optimizes the weight parameters of the inverse filter. In addition, since the inverse filter that suppresses the transmission response has a high computational cost and is not suitable for use in environments with limited computing resources such as real-time embedded devices, the long filter is decomposed into multiple short filters for processing, thereby effectively reducing the computational cost. After the second voice signal is processed by the inverse filter with optimized weight parameters, the third voice signal is obtained, which can effectively suppress the small space transmission response in the spacecraft cabin. After the above three steps of processing, the linear noise in the spacecraft cabin is effectively suppressed, and the obtained third voice signal is used for voice interaction functions such as voice recognition;

[0053] (4) The third voice signal is processed by a nonlinear mapping network to obtain a fourth voice signal and suppress nonlinear noise. The above method can effectively suppress linear noise, but the effect of processing nonlinear noise is not ideal. Therefore, the present invention suppresses nonlinear environmental noise by training a nonlinear mapping network. Like the above method, the noise suppression method examples of the present invention all work in a fixed spacecraft, and can collect a large amount of environmental noise data by recording and other means. According to the prior noise information, the mapping network is trained for nonlinear noise, thereby achieving a better suppression effect. After the third voice signal is processed by the pre-trained nonlinear mapping network, a fourth voice signal is obtained, which has effectively suppressed the nonlinear environmental noise. The obtained fourth voice signal is used for voice communication function.

[0054] Example:

[0055] A noise suppression method for aerospace voice interaction equipment is proposed to solve the pain points of large noise and complex noise environment in the spacecraft cabin, which affects the performance of the voice interaction system. A noise suppression method based on prior information, combined with spatial filtering, post-filtering, inverse filtering, and nonlinear mapping network is proposed. The method is integrated into the algorithm processing control submodule of the voice interaction device, which has the advantages of lightweight and modularization, and achieves a better multi-channel noise suppression effect. The application structure diagram of the present invention is shown in the figure Figure 1 shown.

[0056] An algorithm processing control module for a noise suppression method for aerospace voice interaction equipment is mainly composed of a delayed sum beamforming method based on preset directions and pre-trained spatial filter weight parameters, a post-Wiener filter parameter adaptive correction method based on a priori parameter correction table, a weighted prediction error method based on a priori transmission response, and a nonlinear mapping network method based on pre-training. The overall method flow chart is as follows: Figure 2 shown.

[0057] The flow chart of the delayed sum beamforming method based on preset directions and pre-trained spatial filter weight parameters is as follows: Figure 3 shown.

[0058] In order to effectively save computing resources, this scheme adopts a noise suppression method based on fixed beamforming. Fixed beamforming often uses the delayed summation method, which has low computational overhead and good noise reduction effect. The principle of beamforming technology is to distinguish the speech signal at the desired position from the signal at the interference position to achieve noise suppression. Beamforming is to perform spatial filtering on the received signal after each array element receives the signal, so that the output of the array has spatial directivity and suppresses the influence of noise. The fixed beamforming method means that the weight of the spatial filter is fixed and does not change. It is only related to its own settings and has nothing to do with the signal received by the microphone array. The direction of beamforming depends on the weight of the filter. Its scope of application is mainly the sound field with a relatively certain direction of the signal source, which is consistent with the current noise environment in the spacecraft cabin. The delayed summation beamforming method is a typical fixed beamforming method. It delays and compensates the signals received by different microphones to make the phases of multiple signals the same, and then sums all the signals after weighting, and finally outputs the formed beam signal. In the weighting process, the desired signal is amplified, the noise signal is suppressed, and the reorganized noise reduction signal is obtained.

[0059] Since the simple delay-sum beamforming method is highly dependent on the direction of the signal source, it may be difficult to achieve the ideal noise reduction effect. However, since the noise suppression method examples of the present invention all work in a fixed spacecraft, the direction of the signal source can be preset by a large number of sound sources (including voice signals and noise signals) collected and analyzed in advance. In addition, the weight parameters of the spatial filter represent the accumulated weights of the multi-channel signals collected by the microphone array, which also have a significant impact on the noise suppression effect of the method. Similarly, the prior sound source distribution can be used to determine the optimized weight parameters through a pre-trained spatial filtering model to improve the noise suppression effect of the method. The delay-sum beamforming method can be divided into three parts, namely time delay estimation, time delay compensation and weighted summation. Among them, the present method makes corresponding improvements to these three parts, using the preset optimization of the signal source direction. Time delay estimation, and accurate compensation for time delay, and finally weighted summation of multi-channel signals using the optimized spatial filtering weight parameters obtained by pre-training. Step (1) of the present method is mainly implemented by the following steps:

[0060] Step 1-1: Preset the direction based on the sound sources collected and analyzed in large quantities in advance, and determine the optimized weight parameters through the pre-trained spatial filter model;

[0061] Step 1-2: Estimating the delay according to the above preset direction;

[0062] Step 1-3: Compensate for time delay to align multi-channel signals;

[0063] Step 1-4: Use the optimized spatial filter weight parameters obtained by the above pre-training to perform weighted summation on the multi-channel signals;

[0064] Step 1-5: Output the first speech signal after spatial filtering.

[0065] The commonly used delay-and-sum beamforming method has a strong dependence on the direction of the signal source. The present invention introduces the prior knowledge of the distribution of sound sources in the spacecraft into the method, so as to better determine the directional gain of the spatial filter and enhance the noise suppression effect of the method.

[0066] In addition, the improved beamforming method still requires a large number of microphones to achieve a good noise suppression effect, so the present invention reduces the number of microphones required by a post-Wiener filter. However, in a complex noise environment, the parameters of the Wiener filter obtained by estimation may not be able to achieve an ideal noise reduction effect, especially the low-frequency noise suppression effect is insufficient. The present invention improves the post-Wiener filter and uses a post-Wiener filter parameter adaptive correction method based on a priori parameter correction table to improve the noise suppression effect of the post-Wiener filter.

[0067] like Figure 5 , as in the above method, since each noise reduction algorithm instance works in a fixed spacecraft, a large amount of noisy speech corpus of the scene and the ideal noise reduction corpus obtained by complex technical means can be obtained in advance, and then the corresponding ideal Wiener filter parameters can be estimated. In the present invention, the deviation between the actual estimated filter parameters and the filter parameters under ideal conditions is calculated, and a Wiener filter parameter correction table is formed by discretization, and each table item has a large amount of noise corpus corresponding to it. In the offline training stage, a classifier is trained to map the noise corpus to the corresponding table item, which is called "table lookup classifier" in the present invention. The input is the statistical value of the acoustic representation of several consecutive frames without speech segments, and the output is the table item class. The present invention introduces our prior knowledge of the current spacecraft environmental noise distribution into the noise reduction algorithm through this Wiener filter parameter correction table and table lookup classifier to obtain higher quality noise reduction speech. In the actual working process, after estimating the parameters of the Wiener filter, based on the surrounding continuous speechless corpus, the "lookup classifier" is used to realize adaptive mapping to the specific filter correction parameters in the table, and then the corrected Wiener filter is used to achieve noise suppression.

[0068] The flow chart of the method for adaptively correcting the parameters of the post-Wiener filter based on the prior parameter correction table is as follows: Figure 4 As shown, step (2) is mainly implemented by the following steps:

[0069] Step 2-1: obtain the first speech signal after spatial filtering through the above steps, and use the leading NIS frame silent segment to obtain the average power spectrum and its average amplitude of the noise;

[0070] Step 2-2: Use the endpoint detection method to distinguish between speech frames and silent frames, and use the found silent frames to continuously modify the power spectrum and amplitude spectrum of the estimated noise. Calculate the posterior signal-to-noise ratio and the prior signal-to-noise ratio for the speech frame, so that the transfer function of the frame can be obtained, that is, the expression of the filter parameters;

[0071] Step 2-3: Based on the continuous silent frames, calculate the first-order statistics of their MFCC acoustic representation, and classify them through the table lookup classifier, so as to quickly correspond to a specific table item in the prior parameter correction table, which stores a specific filter parameter correction rule. Based on the rule, the filter parameters are corrected to obtain the corrected Wiener filter;

[0072] Step 2-4: Calculate the output of the Wiener filter to obtain a second speech signal;

[0073] Furthermore, in the third step (step S2-3) of the method for adaptively correcting the parameters of the post-Wiener filter based on the a priori parameter correction table proposed in the present invention, the filter parameter correction table constructed in advance is used to adaptively determine the filter parameter correction value based on the statistical characteristics of the noise, thereby obtaining the corrected post-Wiener filter. The above-mentioned post-Wiener filter parameter correction process is as follows: Figure 5 As shown, it is mainly implemented by the following steps:

[0074] Step 2-3-1: Calculate the MFCC cepstrum features of the continuous silent frames around the current processing spectrum, and calculate their mean and variance;

[0075] Step 2-3-2: The first-order statistics of these silent frames are fed into the "lookup classifier" as input for classification decision, and the determined category number is the table entry number in the corresponding filter parameter correction table;

[0076] Step 2-3-3: Take out the filter parameter correction value stored in the table entry and correct the filter.

[0077] In addition, for the specific enclosed space and transmission medium in the spacecraft cabin, this scheme adopts a weighted prediction error method based on a priori transmission response. Combining the statistical model of speech time variation and the multi-channel linear prediction filter, an efficient method for suppressing speech transmission response can be generated, namely the weighted prediction error method. This method deals with the transmission response problem at the signal level and is based on delayed linear prediction. The weighted prediction error method first obtains a frequency-dependent linear prediction filter through a series of historical frame information. Then, the enhanced speech signal is obtained in the subband domain by taking the difference between the original response signal and the filtered signal.

[0078] Linear prediction methods have shown good results in estimating the inverse filter of an unknown system. When linear prediction analysis is applied to transmission-responsive speech, the influence of the transmission response is mainly present in the prediction residual, especially in the case of a multi-microphone system. In such a system, the accuracy of estimating the linear prediction coefficients of clean speech from the transmission response observations is improved. Suppressing the transmission response is achieved by processing the linear prediction residual signal and then synthesizing a speech signal with reduced transmission response from the output of the filter using the linear prediction coefficients obtained from the transmission-responsive speech.

[0079] The weighted prediction error method estimates the delayed transmission response component based on a multi-channel linear predictor. Normally, the prediction filter in each short-time Fourier transform subband must be long enough to accurately estimate the delayed transmission response component. Therefore, the computational cost of the weighted prediction error method is very high, which makes it difficult to implement in devices with limited computing resources such as real-time embedded systems. Since the Kronecker product can decompose a long filter into many short filters for processing, the present invention introduces the Kronecker integral decomposition to improve the traditional weighted prediction error method for suppressing transmission response, so that the computational complexity of the method is significantly reduced. The present invention proposes to decompose the multi-channel linear prediction filter into three sub-filters, thereby converting a high-dimensional problem into a low-dimensional problem for processing.

[0080] At the same time, since the noise reduction algorithm example of the present invention works in a fixed spacecraft, the small space transmission response in the cabin can be analyzed through a large number of voice signals collected in advance, and then the parameters of the optimized linear prediction filter can be estimated.

[0081] In step (3), the flow chart of the weighted prediction error method based on the priori transmission response is as follows: Figure 6 As shown, it is mainly implemented by the following steps:

[0082] Step 3-1: After the above steps, a second speech signal after post-filtering is obtained, the signal includes a spatial transmission response, and a desired signal is modeled using a time-varying Gaussian model;

[0083] Step 3-2: Use the maximum likelihood criterion to estimate the unknown parameters of the multi-channel linear prediction estimator to obtain the probability density function of the expected signal;

[0084] Step 3-3: Estimate the above unknown parameters so that the probability density of the desired signal at the current frequency point is as large as possible;

[0085] Step 3-4: Introduce the Kronecker product to decompose the prediction filter into three sub-filters, and use the method in the above steps to iteratively calculate the parameters of the three sub-filters;

[0086] Step 3-5: Introduce the information of prior spatial transmission response and optimize the above filter parameters;

[0087] Step 3-6: The output of the second speech signal passing through the above filter is used as the third speech signal.

[0088] The third voice signal obtained by the above method effectively suppresses the small space transmission response in the spacecraft cabin, has a good suppression effect on linear noise, and is suitable as an input to realize voice interaction functions such as voice recognition. However, for nonlinear noise, it is still necessary to use prior noise information for further processing to suppress environmental noise and obtain better voice communication effects.

[0089] Similar to the above method, since each denoising algorithm instance works in a fixed spacecraft, a large amount of environmental noise of the scene can be acquired in advance, and the nonlinear part thereof can be effectively suppressed by using a pre-trained nonlinear mapping network.

[0090] In step (4), the flowchart of the pre-trained nonlinear mapping network method is as follows: Figure 7 As shown, it is mainly implemented by the following steps:

[0091] Step 4-1: Obtain a third speech signal after inverse filtering through the above steps, and input the signal into a nonlinear mapping network;

[0092] Step 4-2: Using the prior environmental noise information, train the nonlinear mapping network to suppress the nonlinear part of the environmental noise;

[0093] Step 4-3: Obtaining an output through the nonlinear mapping network to obtain a fourth speech signal;

[0094] The fourth voice signal obtained through the above steps can achieve a good ambient sound noise reduction effect, and thus can be used as input for voice communication.

[0095] The present invention has been described in detail above in conjunction with specific implementations and exemplary examples, but these descriptions cannot be understood as limiting the present invention. Those skilled in the art understand that, without departing from the spirit and scope of the present invention, a variety of equivalent substitutions, modifications or improvements may be made to the technical solution of the present invention and its implementation methods, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be subject to the attached claims.

[0096] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A noise suppression method for aerospace voice interaction equipment, characterized in that: include: S1 processes the original speech signal using a spatial filter to obtain a first speech signal with spatial directivity; S2 performs post-filtering processing on the first speech signal to suppress low-frequency noise of the first speech signal to obtain a second speech signal; S3 performs inverse filtering on the second voice signal to suppress a small space transmission response of the second voice signal to obtain a third voice signal; S4 utilizes a nonlinear mapping network to suppress nonlinear noise in the third voice signal to obtain a fourth voice signal suitable for aerospace voice communication.

2. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: The method of step S1 comprises: S1.1 presets the direction according to the previously collected sound source, and determines the optimized spatial filter weight parameters through the pre-trained spatial filter model; S1.2 estimates the delay of the acquired original multi-channel speech signal according to a preset direction to obtain a multi-channel delay signal; S1.3 performs delay compensation on multi-channel delay signals to align the multi-channel delay signals; S1.4 uses the optimized spatial filtering weight parameters to perform weighted summation on the aligned multi-channel time-delayed signals to obtain a single-channel signal, that is, a first speech signal with spatial directivity.

3. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: The method of step S2 includes: Obtain noisy speech corpus in advance, and obtain the denoised corpus under ideal conditions; Obtaining corresponding ideal post-filter parameters according to the noisy speech corpus and the ideal denoised corpus; The first speech signal is processed using the post-filter parameters under ideal conditions to obtain a second speech signal.

4. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: In step S3, the filter used when inverse filtering is performed on the second speech signal includes a plurality of short filters.

5. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: In step S4, the nonlinear mapping network is trained using the priori noise information obtained in advance, and the nonlinear noise in the third speech signal is suppressed using the trained nonlinear mapping network.

6. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: The method of step S2 includes: S2.1 constructs a Wiener filter parameter correction table and a table lookup classifier; The Wiener filter parameter correction table is used to describe the correspondence between the noise corpus and the correction rules; The table lookup classifier is used to describe the correspondence between the statistical values ​​of the acoustic representation of a number of consecutive frames without speech and the correction rules; S2.2 calculating the posterior signal-to-noise ratio and the prior signal-to-noise ratio using the speaking frame in the first speech signal to obtain the posterior signal-to-noise ratio and the prior signal-to-noise ratio of the first speech signal in the frame, and obtaining the transfer function of the first speech signal in the frame according to the posterior signal-to-noise ratio and the prior signal-to-noise ratio, that is, the filter parameters of the first speech signal in the frame; S2.3 calculating the statistical value of the acoustic representation of a plurality of consecutive speechless segments in the first speech signal, and obtaining the corresponding correction rule according to the table lookup classifier obtained in step S2.1; S2.4 Based on the Wiener filter parameter correction table obtained in step S2.1, the filter parameters of the first speech signal obtained in step S2.2 are corrected using the correction rule obtained in step S2.3 to obtain corrected filter parameters; S2.5 performs post-filtering processing on the first speech signal using the modified filter parameters obtained in step S2.4 to obtain a second speech signal.

7. A noise suppression method for aerospace voice interaction equipment according to claim 6, characterized in that: In step S2.1, the statistical values ​​of the acoustic representation of a plurality of consecutive frames without speech include the mean and variance of the MFCC cepstrum features of the silent frames.

8. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: In step S3, firstly, a frequency-dependent linear prediction filter is obtained through a series of historical frame information, and then, the enhanced speech signal is obtained in the subband domain through the difference between the original response signal and the filtered signal.

9. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: In step S3, Kronecker integral decomposition is introduced to decompose the multi-channel linear prediction filter into three sub-filters.

10. The noise suppression method for aerospace voice interaction equipment according to claim 1, characterized in that: In step S1.2, the method for determining the preset direction includes: The direction of the maximum output power pointed by the multi-channel speech signal beam collected by the microphone array is used as the preset direction.

Images