Adaptive intelligent noise reduction method for transmission of audio system

Abstract

The invention relates to the technical field of digital information transmission, in particular to an audio system transmission adaptive intelligent noise reduction method, which comprises the following steps: firstly, acquiring transmission channel state information, and designing a dynamic pilot frequency to extract key channel parameters; a double-ring processing model is adopted, an inner ring carries out time delay compensation through an LMS algorithm, an outer ring estimates channel gain through Kalman filtering, and channel impulse response is updated to optimize channel state recovery. Then, extracting audio features, performing voice noise reduction in combination with the improved NMF model and the noise dictionary, and optimizing the noise dictionary in real time through a dynamic dictionary mechanism; and the noise-reduced voice is subjected to clock synchronization by adopting a PTP protocol, and an NACK retransmission request is executed or decoding is performed through FEC according to a packet loss rate detection result, so that the transmission accuracy and reliability of the audio system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital information transmission, and in particular to an audio system transmission adaptive intelligent noise reduction method. Background Art

[0002] In the process of data transmission and signal processing, especially in large-scale space application scenarios, the stability, anti-interference ability and end-to-end synchronization processing delay of audio data in network transmission have a significant impact on the quality of audio communication. Therefore, it is of great practical significance to develop an audio data transmission method based on adaptive signal processing and intelligent noise reduction.

[0003] In large-scale spatial environments, existing adaptive transmission schemes are usually based on preset channel state assumptions, which are difficult to deal with sudden changes in acoustic parameters caused by personnel movement, and cannot effectively deal with multipath time-varying effects caused by building structures. In addition, in distributed audio data transmission systems, there are multiple sources of noise interference, such as air conditioning fans, electromagnetic interference, keyboard tapping, and network jitter. The spectral characteristics of these interferences may overlap greatly. When multiple noise sources interfere with each other, these adaptive transmission schemes are difficult to effectively model the spatial distribution of noise, thereby affecting the accuracy of noise reduction, and further leading to an increase in the bit error rate of data transmission. Finally, in a multi-device collaborative audio data transmission network, it is usually necessary to synchronize and fuse multiple audio data streams. However, these adaptive transmission schemes adopt a centralized data processing architecture and rely on a central server for data aggregation and signal synchronization, which increases the delay of audio data transmission. When the network load is high, problems such as packet loss and computing resource bottlenecks may occur, thereby affecting the real-time performance of audio data and the synchronous noise reduction effect.

[0004] In order to improve the accuracy of audio system transmission, an adaptive intelligent noise reduction method for audio system transmission is proposed. Summary of the invention

[0005] The purpose of the present invention is to provide an adaptive intelligent noise reduction method for audio system transmission, by acquiring transmission channel state information and designing dynamic pilots to extract key channel parameters. A dual-loop processing model is adopted, in which the inner loop performs delay compensation through the LMS algorithm, and the outer loop estimates the channel gain using Kalman filtering, and updates the channel impulse response to optimize channel state recovery. Then, the audio features are extracted, and speech noise reduction is performed in combination with the improved NMF model and the noise dictionary, and the noise dictionary is optimized in real time through the dynamic dictionary mechanism. The denoised speech uses the PTP protocol for clock synchronization, and executes a NACK retransmission request or performs FEC decoding according to the packet loss rate detection result, thereby improving the accuracy and reliability of the audio system transmission.

[0006] To achieve the above object, the present invention provides the following technical solutions: An audio system transmission adaptive intelligent noise reduction method, comprising: Obtaining transmission channel status information; Design a dynamic pilot to extract the pilot part according to the channel parameters and environment parameters of the transmission channel state information; Establishing a dual-loop processing model, including: inner loop processing, for aligning the pilot part with the time domain of the local reference sequence using the LMS algorithm, and outputting the delay compensation; outer loop processing, for estimating the channel gain using the Kalman filter; using the delay compensation and the channel gain to update the channel impulse response and restore the transmission channel state information; Extracting audio features of the transmission channel state information, decomposing the noise reduction speech and the noise floor using an improved NMF model and a noise dictionary for the audio features; and updating the noise dictionary using a dynamic dictionary mechanism; The noise reduction speech is clock-synchronized using the PTP protocol and the packet loss rate is detected. If the packet loss rate is greater than a preset condition, a NACK retransmission request is sent; otherwise, the noise reduction speech is decoded using FEC according to a dynamic redundancy mechanism.

[0007] Furthermore, the transmission channel state information includes: The channel parameters include multi-channel speech signals collected by a microphone array; Wireless network channel parameters, including network bandwidth, packet loss rate and delay jitter; The environmental parameters include the three-dimensional size of the space, the location of the sound source, the material of the reflective surface and the number of people.

[0008] Furthermore, the design process of the dynamic pilot includes: The pilot part is defined using a dynamic adjustment strategy, which is expressed as: ; in, is the pilot part, is the pilot amplitude adjusted according to the environmental parameters, is the cosine function, is the pilot frequency, is the time variable; Extract the pilot part according to the pilot part , expressed as: ; in, is the channel impulse response, is additive noise, is the convolution operation, is a discrete time index.

[0009] Furthermore, the audio feature extraction process includes: Using a Gammatone filter bank to perform multi-channel frequency band decomposition on the transmission channel state information to obtain an audio sub-band signal; Calculating a spatial covariance matrix for the audio subband signal for analyzing spatial distribution characteristics; MFCC feature extraction is performed on the audio subband signal, and the spectrum time series feature is obtained by combining the dynamic difference coefficient.

[0010] Furthermore, the optimization objective function of the improved NMF model is include: ; in, is the audio feature matrix, is the basis matrix, is the weight matrix, is the time domain continuity constraint term, is the basis sparsity constraint term, and To adjust the weights, is the Frobenius norm, is the weight matrix of the kth frame, is the weight matrix of the k-1th frame.

[0011] Furthermore, the noise dictionary construction process includes: Collect noise samples; For each type of the noise samples, extract the noise floor using the improved NMF model; For the noise basis, a basis matrix that satisfies an energy concentration threshold and a time continuity threshold is screened to generate the noise dictionary.

[0012] Furthermore, the updating process of the noise dictionary includes: Calculate the matching error between the noise floor and the noise dictionary , expressed as: ; in, is the audio feature, is the existing noise floor, is the weight matrix corresponding to the existing noise floor, is the Frobenius norm; If it is detected that the matching error exceeds the set threshold and lasts for more than N seconds, a new noise floor is calculated The orthogonal projection residual , expressed as: ; in, is the existing noise floor, for of pseudo-rebellion; like If the residual error is greater than the preset threshold, The noise dictionary is added and optimized through atom pruning and Gram-Schmidt orthogonalization.

[0013] Furthermore, the dynamic redundancy mechanism includes: adjusting redundancy according to the packet loss rate; According to the redundancy, the FEC is used to optimize the audio transmission quality for encoding parameters under different network conditions.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention uses a dynamic pilot extraction mechanism to adaptively extract the pilot part according to the time-varying characteristics of the channel state information, thereby improving the accuracy of channel estimation. In addition, by constructing a dual-loop processing model, the inner loop processing uses the LMS algorithm for time domain alignment to compensate for the delay error, and the outer loop processing uses the Kalman filter to estimate the channel gain in real time, and finally combines the delay compensation and gain information to optimize the channel impulse response, thereby effectively improving the accuracy of channel state recovery and enhancing the adaptability and robustness of the audio system in complex environments.

[0015] 2. The present invention improves the accuracy of audio feature extraction through Gammatone filter, spatial covariance matrix and MFCC feature extraction, and combines the improved NMF model with the noise dictionary for speech decomposition and noise reduction, thereby achieving efficient audio signal enhancement. In addition, the dynamic dictionary update mechanism is used to detect noise changes in real time and dynamically expand the noise dictionary to ensure the adaptive noise reduction capability in complex noise environments, thereby improving the clarity and separability of audio signals and enhancing the accuracy of audio recognition.

[0016] 3. The present invention integrates PTP protocol synchronization and adaptive packet loss recovery mechanism. In the process of noise-reduced voice data transmission, high-precision clock synchronization is achieved through the PTP protocol, which effectively reduces the transmission delay error. At the same time, the packet loss rate is monitored in real time. If the packet loss rate exceeds the threshold, a NACK request is triggered for data retransmission; otherwise, FEC error correction decoding is used based on the dynamic redundancy mechanism to improve the stability of voice data transmission and enhance the communication reliability and noise reduction accuracy of the audio system in a complex network environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a flow chart of an audio system transmission adaptive intelligent noise reduction method provided by the present invention; Figure 2 A schematic diagram of the process of decomposing noise-reduced speech and noise floor provided by the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] See also Figure 1 to Figure 2 The present invention provides an audio system transmission adaptive intelligent noise reduction method, and the technical solution is as follows: In large-scale spatial and distributed audio transmission systems, audio signal transmission faces many challenges. For example, dynamic acoustic environments and multipath effects may make it difficult for traditional solutions to adapt, resulting in unstable signals. In addition, multi-source noise interference is difficult to effectively distinguish and suppress, affecting transmission quality. Finally, centralized processing architectures are prone to bottlenecks under high loads, increasing latency and making it difficult to meet personalized noise reduction needs. Existing noise reduction solutions lack adaptive optimization capabilities and are difficult to cope with complex and changing environments and channel conditions.

[0020] In order to improve the accuracy of audio system transmission in a large space, the first embodiment is as follows: Figure 1 A schematic diagram of a flow chart of an audio system transmission adaptive intelligent noise reduction method provided by the present invention.

[0021] like Figure 1 As shown, an audio system transmission adaptive intelligent noise reduction method includes: Step 1: Obtain transmission channel status information; Furthermore, the transmission channel state information includes: Channel parameters, including multi-channel speech signals collected by microphone arrays; Wireless network channel parameters, including network bandwidth and packet loss rate, are used to adjust audio coding and error correction mechanisms to ensure data reliability; delay jitter is used to monitor network transmission jitter and optimize timing synchronization strategies; Environmental parameters include the three-dimensional dimensions of the space, the location of the sound source, the material of the reflective surface, and the number of people.

[0022] In addition, it can also include device status information, such as microphone gain and frequency response curve, for adaptive adjustment of signal processing parameters. By obtaining this transmission channel status information, the audio signal processing strategy can be adjusted in real time, the data transmission quality can be optimized, and the anti-interference ability can be enhanced, thereby improving the accuracy of audio system transmission.

[0023] Step 2: Design a dynamic pilot to extract the pilot part according to the channel parameters and environment parameters of the transmission channel state information.

[0024] Furthermore, the design process of the dynamic pilot includes: The pilot part is defined using a dynamic adjustment strategy, which is expressed as: ; in, is the pilot part, is the pilot amplitude adjusted according to the environmental parameters, is the cosine function, is the pilot frequency, is the time variable; Extract the pilot part according to the pilot part , expressed as: ; in, is the channel impulse response, is additive noise, is the convolution operation, is a discrete time index.

[0025] By dynamically adjusting the pilot transmission interval when the density of people changes, it is possible to maintain efficient channel tracking capabilities, capture channel changes more timely, and improve the accuracy of audio transmission.

[0026] Step 3: Establish a dual-loop processing model, including: inner loop processing, used to use the LMS (Least Mean Squares) algorithm to align the pilot part with the time domain of the local reference sequence and output delay compensation; outer loop processing, used to estimate the channel gain using a Kalman filter; use the delay compensation and the channel gain to update the channel impulse response and restore the transmission channel state information.

[0027] Specifically, in this embodiment, in the inner loop processing, the local reference sequence is used for delay compensation. The transmitter generates a fixed pilot part in advance and inserts it into the audio data stream during the transmission process. The receiver stores the same local reference sequence for comparing the received pilot part and calculating the delay error. The received signal is adjusted by the minimum mean square error algorithm to align it with the local reference sequence on the time axis, thereby eliminating the delay drift during the transmission process. In the outer loop processing, the Kalman filter can estimate the channel gain in real time to reflect the dynamic changes of the channel and optimize the signal quality under the multipath effect. Next, the joint processing is combined with the delay compensation. and channel gain Update channel impulse response , thereby reducing multipath interference and improving the accuracy of signal recovery.

[0028] Step 4: Figure 2 This is a schematic diagram of the process of decomposing noise-reduced speech and noise floor provided by the present invention. Figure 2 As shown, the spatial distribution characteristics and spectral time series characteristics of the transmission channel state information are extracted and fused to form a joint feature matrix to obtain audio features. The audio features are decomposed into de-noised speech and noise floor using an improved NMF (Non-negative Matrix Factorization) model and a noise dictionary; and the noise dictionary is updated using a dynamic dictionary mechanism.

[0029] Furthermore, the audio feature extraction process includes: The transmission channel state information is decomposed into multi-channel frequency bands using a Gammatone filter bank to obtain an audio sub-band signal, which is expressed as: ; in, is the audio subband signal, To transmit channel state information, The center frequency is Gammatone filter; This embodiment uses 32 filters with a center frequency range of 50 Hz to 8000 Hz.

[0030] Calculate the spatial covariance matrix of the audio subband signal , used to analyze spatial distribution characteristics, expressed as: ; in, is the time window length, is the multi-channel audio subband signal matrix of the kth frame, is the matrix conjugate transpose; Among them, this embodiment calculates the covariance matrix within a 50ms time window, extracts the maximum eigenvalue and its corresponding eigenvector, and uses them to enhance the target speech signal.

[0031] The audio subband signal is subjected to MFCC (Mel-Frequency Cepstral Coefficients) feature extraction, and the spectrum time series feature is obtained by combining the dynamic differential coefficient. Specifically, the method includes: using short-time Fourier transform to calculate the short-time power spectrum of the audio subband signal, applying the Mel filter bank to obtain the Mel spectrum, taking the logarithm of the Mel spectrum, and then performing discrete cosine transform to obtain MFCC, and calculating the dynamic differential coefficient according to the MFCC to form the spectrum time series feature.

[0032] Gammatone filtering is used to enhance the key frequency band signals of the transmission channel state information, effectively reducing noise interference. At the same time, the covariance matrix is ​​calculated to analyze the sound source distribution and improve the ability to extract spatial information. On this basis, MFCC and its dynamic difference are extracted to further enhance the robustness of speech features, thereby improving the accuracy of audio transmission.

[0033] Furthermore, the optimization objective function of the improved NMF model includes: ; in, is the audio feature matrix, each column represents a time frame, and each row represents a feature dimension. is the basis matrix, each column represents a fundamental frequency mode, is a weight matrix, each column represents the activation degree of each basis at each time frame, It is a time domain continuity constraint item, which ensures the smooth time variation of the speech signal and reduces the mutation caused by noise. is the basis sparsity constraint term, which ensures that the speech and noise basis do not overlap as much as possible and improves the separability of the speech. and To adjust the weights, is the Frobenius norm.

[0034] in, is the weight matrix of the kth frame, is the weight matrix of the k-1th frame.

[0035] Specifically, first input the pre-trained noise dictionary , the basis matrix of the target speech (obtained from a clean speech library) and audio features. Next, initialize the basis matrix and the weight matrix , is a random value. Then, for the optimization objective function and Use gradient descent to update the parameters. The weight matrix obtained by decomposition Determine which bases belong to speech and which belong to noise, and keep The relevant part is removed, which is related to the noise floor. , and obtain the audio signal after noise reduction.

[0036] The improved NMF model introduces time domain continuity constraints to reduce speech signal mutations and improve speech smoothness. At the same time, the base sparsity constraint enhances the distinction between speech and noise base and prevents noise from mixing into speech. In addition, the use of noise dictionaries enables the method to adapt to a variety of background noises and improves the accuracy of audio transmission.

[0037] Furthermore, the noise dictionary construction process includes: Collect noise samples; For each type of the noise samples, extract the noise floor using the improved NMF model; For the noise basis, a basis matrix that satisfies an energy concentration threshold and a time continuity threshold is screened to generate the noise dictionary.

[0038] Specifically, in this example, common noise categories such as keyboard tapping and air conditioner operation are first extracted from the public dataset, the sampling rate is set to 16kHz, and the length is truncated to 10s to ensure consistency. The noise samples are converted into time-frequency features, and the noise floor is extracted using the improved NMF model. , , and for different noise floor modes.

[0039] Calculate the energy concentration, expressed as: ; in, is the energy concentration of the ath basis, is the a-th row and b-th column element of the basis matrix, is the b-th row and a-th column element of the weight matrix; Set the energy concentration threshold. If If the energy concentration is greater than the threshold, the basis is retained, otherwise it is discarded.

[0040] Compute time continuity, expressed as: ; in, is the time continuity of the a-th basis, is the activation value of the weight matrix at b; Set the time continuity threshold. If If it is less than the time continuity threshold, the basis is retained, otherwise it is discarded.

[0041] The final noise floor is determined by taking the intersection of the two thresholds. A noise dictionary is constructed and pre-trained, and known noise patterns are directly used in noise reduction. The pre-trained dictionary only requires sparse representation matching for denoising, which has lower computational overhead than online training and provides an accurate noise floor for audio system denoising.

[0042] Furthermore, the updating process of the noise dictionary includes: Calculate the matching error between the noise floor and the noise dictionary , expressed as: ; in, is the audio feature, is the existing noise floor, is the weight matrix corresponding to the existing noise floor, is the Frobenius norm; If it is detected that the matching error exceeds the set threshold and lasts for more than N seconds, a new noise floor is calculated The orthogonal projection residual , expressed as: ; in, is the existing noise floor, for of pseudo-rebellion; like If the residual error is greater than the preset threshold, The noise dictionary is added and optimized through atom pruning and Gram-Schmidt orthogonalization.

[0043] Atomic pruning can remove redundant or weakly activated bases to prevent the dictionary from being too large, while Gram-Schmidt orthogonalization ensures that the dictionary bases are approximately orthogonal, enhancing the noise separation effect. The updated noise dictionary can more accurately represent new noise, improve the quality of audio system extraction, and thus ensure the noise reduction effect in large space environments.

[0044] Step 5: Use the PTP (Precision Time Protocol) protocol to perform clock synchronization marking on the noise reduction speech and detect the packet loss rate. If the packet loss rate is greater than a preset condition, send a NACK (Negative Acknowledgment) retransmission request; otherwise, use FEC (Forward Error Correction) to decode the noise reduction speech according to the dynamic redundancy mechanism.

[0045] Furthermore, the dynamic redundancy mechanism includes: adjusting redundancy according to the packet loss rate; According to the redundancy, the FEC is used to optimize the audio transmission quality for encoding parameters under different network conditions.

[0046] Specifically, in order to ensure the timing consistency and data integrity of audio data, a timestamp based on the PTP protocol is added to each data packet during the transmission of audio data after noise reduction to ensure time synchronization of all receiving devices and avoid audio distortion caused by disordered data packets. The receiving end detects the packet loss rate. If the packet loss rate is greater than the preset threshold, a NACK request is sent to require the sending end to retransmit the lost voice data packet. If the packet loss rate is not greater than the preset threshold, FEC is used for decoding according to the dynamic redundancy mechanism. For example, assuming that the current network jitter is small, the packet loss rate is 1%, which is lower than the preset threshold of 5%, and the redundancy is set to 10%, some lost data packets can be recovered. Assuming that the current network jitter is large, the packet loss rate is 4%, which is still less than the preset threshold. The FEC mechanism adjusts the redundancy to 20% to enhance the error correction capability.

[0047] Compared with fixed redundancy FEC, the dynamic redundancy mechanism can adaptively adjust the redundancy according to network conditions, which not only improves transmission efficiency, but also enhances the accuracy of audio transmission, providing users with a more reliable audio experience.

[0048] First, the pilot is extracted in real time through dynamic pilot design, which improves the accuracy of channel information acquisition. Then, a dual-loop processing model is adopted. The inner loop uses the LMS algorithm to align the signal to compensate for the delay, which reduces the audio distortion. The outer loop uses Kalman filtering to estimate the channel gain and enhances the data recovery capability. The channel model is dynamically adjusted by combining delay compensation and channel gain to adapt to complex environments. Then, an improved NMF denoising model is designed, which uses time domain continuity and base sparsity constraints to improve the ability to separate speech noise. Combined with the dynamic dictionary update mechanism, new bases are automatically added when the environmental noise changes, which can stabilize the noise reduction effect. Finally, during transmission, the packet loss detection and retransmission mechanism is used, and the FEC error correction strength is adjusted with dynamic redundancy to balance bandwidth and error correction, thereby improving the accuracy of audio transmission.

[0049] Embodiment 2: In modern smart conference rooms, since conference rooms are usually large enclosed spaces, their audio transmission systems are easily affected by factors such as multipath effects, environmental noise, and personnel movement, which leads to dynamic changes in the transmission channel. Traditional noise reduction and signal processing methods are difficult to meet actual needs. Taking a technology company's conference room as an example, an audio system transmission adaptive intelligent noise reduction method is proposed to achieve efficient noise reduction and channel adaptive processing to ensure the accuracy of conference audio, including: Step 1: Obtain transmission channel status information; Furthermore, in this embodiment, the transmission channel state information includes: Channel parameters include: multi-channel voice signals collected by microphone arrays, combined with multipath effect parameters (signal attenuation and delay), echo path characteristics (room echo path model calculated based on echo cancellation algorithm), and signal-to-noise ratio information (used for dynamic gain control) to analyze the impact of the environment on signal propagation.

[0050] Wireless network channel parameters, including: network bandwidth and packet loss rate, are used to adjust audio coding and error correction mechanisms to ensure data reliability; delay jitter, the arrival time difference of adjacent data packets is calculated for the timestamps of the received audio data packets, and the standard deviation is calculated using a sliding window (window size 1s) to estimate the delay jitter range, thereby optimizing the timing synchronization strategy.

[0051] Environmental parameters include: three-dimensional space dimensions, which are set by the structural parameters of the conference room. The location of the sound source, which uses the time difference of arrival (TDOA) algorithm to calculate the coordinates of the sound source and determine the location of the sound source; the material of the reflective surface, which is determined based on the material properties of different areas of the conference room in combination with the pre-stored reflection coefficient table; and the number of people, which uses the YOLOv5 model to count the number of people in real time.

[0052] Step 2: Design a dynamic pilot to extract the pilot part based on the channel parameters and environmental parameters of the transmission channel state information. Because in the conference room environment, sound propagation is affected by the multipath effect, causing the signal to reach the microphone after propagating through different paths, resulting in phase shift and amplitude attenuation. These distortions need to be compensated to improve audio quality.

[0053] Furthermore, the design process of the dynamic pilot includes: The pilot part is defined using a dynamic adjustment strategy, using a periodic known signal, which is expressed as: ; in, is the pilot part, is the pilot amplitude adjusted according to the environmental parameters, is the cosine function, is the pilot frequency, set to 500, is the time variable; Extract the pilot part according to the pilot part , expressed as: ; in, is the channel impulse response, is additive noise, is the convolution operation, is a discrete time index.

[0054] Specifically, the size of the conference room is 18m×12m×4m, the sound source is 6m away from the speaker and the microphone, the reflection surface is the glass curtain wall (reflection coefficient 0.85), and the other walls are concrete. The number of people is 15 at the beginning, and it increases to 25 during the meeting. Based on these data, the calculated data is shown in Table 1. The density of people From 0.07 people / square meter to 0.12 people / square meter. When there are 15 people, 1.7, 25 people, The pilot part is set to 2.2 to ensure that the pilot part can still be detected when noise and reflection are enhanced. The pilot transmission interval threshold is set to 0.1, so the pilot is extracted at intervals of 50ms and 30ms respectively. The transmission interval is shortened from 50ms to 30ms, and the channel status is updated more frequently. Through dynamic adjustment, the pilot can more accurately estimate the phase offset and amplitude attenuation caused by multipath effects and noise, thereby improving the accuracy of the audio.

[0055] Table 1 Dynamic pilot data

[0056] Step three: Establish a dual-loop processing model, including: inner loop processing, used to use the LMS algorithm to align the pilot part with the time domain of the local reference sequence and output delay compensation; outer loop processing, used to estimate the channel gain using the Kalman filter; use the delay compensation and the channel gain to update the channel impulse response and restore the transmission channel state information.

[0057] Specifically, as shown in Table 2, the test conditions are that the air conditioner is continuously turned on, there is sudden keyboard tapping, and two people are walking crosswise at a speed of about 1m / s. The single LMS algorithm is used to compare with the dual-loop processing model of the present invention. Table 2 Effect comparison

[0058] Step 4: extract the audio features of the transmission channel state information, use the improved NMF model and noise dictionary to decompose the noise reduction speech and the noise floor of the audio features; and use the dynamic dictionary mechanism to update the noise dictionary.

[0059] Step 5: Use the PTP protocol to perform clock synchronization marking on the noise reduction speech and detect the packet loss rate. If the packet loss rate is greater than a preset condition, send a NACK retransmission request; otherwise, use FEC to decode the noise reduction speech according to the dynamic redundancy mechanism.

[0060] Specifically, the deployment plan for a company's conference room is shown in Table 3. For example, under the 5G network, the packet loss rate is low, so 10% redundant coding is adopted and a packet length of 1KB is used to optimize transmission efficiency. Under this configuration, 3% random packet loss can be effectively repaired to ensure the stability of audio transmission. Compared with the traditional fixed redundancy solution, the average decoding delay of the present invention is about 12ms, and the decoding delay of the traditional fixed redundancy solution is about 16 seconds. The present invention improves the decoding speed and the accuracy of audio system transmission.

[0061] Table 3 Conference room deployment plan

[0062] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An audio system transmission adaptive intelligent noise reduction method, characterized in that: include: Obtaining transmission channel status information; Design a dynamic pilot to extract the pilot part according to the channel parameters and environment parameters of the transmission channel state information; Establishing a dual-loop processing model, including: inner loop processing, for aligning the pilot part with the time domain of the local reference sequence using the LMS algorithm, and outputting the delay compensation; outer loop processing, for estimating the channel gain using the Kalman filter; using the delay compensation and the channel gain to update the channel impulse response and restore the transmission channel state information; Extracting audio features of the transmission channel state information, decomposing the noise reduction speech and the noise floor using an improved NMF model and a noise dictionary for the audio features; and updating the noise dictionary using a dynamic dictionary mechanism; The noise reduction speech is clock-synchronized using the PTP protocol and the packet loss rate is detected. If the packet loss rate is greater than a preset condition, a NACK retransmission request is sent; otherwise, the noise reduction speech is decoded using FEC according to a dynamic redundancy mechanism.

2. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The transmission channel state information includes: The channel parameters include multi-channel speech signals collected by a microphone array; Wireless network channel parameters, including network bandwidth, packet loss rate and delay jitter; The environmental parameters include the three-dimensional size of the space, the location of the sound source, the material of the reflective surface and the number of people.

3. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The design process of the dynamic pilot includes: The pilot part is defined using a dynamic adjustment strategy, which is expressed as: ； in, is the pilot part, is the pilot amplitude adjusted according to the environmental parameters, is the cosine function, is the pilot frequency, is the time variable; Extract the pilot part according to the pilot part , expressed as: ； in, is the channel impulse response, is additive noise, is the convolution operation, is a discrete time index.

4. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The audio feature extraction process includes: Using a Gammatone filter bank to perform multi-channel frequency band decomposition on the transmission channel state information to obtain an audio sub-band signal; Calculating a spatial covariance matrix for the audio subband signal for analyzing spatial distribution characteristics; MFCC feature extraction is performed on the audio subband signal, and the spectrum time series feature is obtained by combining the dynamic difference coefficient.

5. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The optimization objective function of the improved NMF model is include: ； in, is the audio feature matrix, is the basis matrix, is the weight matrix, is the time domain continuity constraint term, is the basis sparsity constraint term, and To adjust the weights, is the Frobenius norm, is the weight matrix of the kth frame, is the weight matrix of the k-1th frame.

6. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The process of constructing the noise dictionary includes: Collect noise samples; For each type of the noise samples, extract the noise floor using the improved NMF model; For the noise basis, a basis matrix that satisfies an energy concentration threshold and a time continuity threshold is screened to generate the noise dictionary.

7. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The updating process of the noise dictionary includes: Calculate the matching error between the noise floor and the noise dictionary , expressed as: ； in, is the audio feature, is the existing noise floor, is the weight matrix corresponding to the existing noise floor, is the Frobenius norm; If it is detected that the matching error exceeds the set threshold and lasts for more than N seconds, a new noise floor is calculated The orthogonal projection residual , expressed as: ； in, is the existing noise floor, for of pseudo-rebellion; like If it is greater than the preset residual threshold, The noise dictionary is added and optimized through atom pruning and Gram-Schmidt orthogonalization.

8. The method for adaptive intelligent noise reduction in audio system transmission according to claim 1, characterized in that: The dynamic redundancy mechanism includes: adjusting redundancy according to the packet loss rate; According to the redundancy, the FEC is used to optimize the audio transmission quality for encoding parameters under different network conditions.

Images