Conference sound amplification system based on AI intelligent algorithm and 360-degree omnidirectional noise reduction

Abstract

The invention discloses a conference sound reinforcement system based on an AI intelligent algorithm and 360-degree omni-directional noise reduction, and relates to the technical field of audio signal processing, the conference sound reinforcement system comprises a conference management center, the conference management center is in communication connection with the following modules: a multi-sound-source sensing module used for capturing 360-degree omni-directional sound field information in a conference environment and constructing a sound source space distribution model; according to the invention, the omnidirectional microphone array unit covers all directions of a conference space, synchronously collects audio data streams, eliminates the limitation of a conventional unidirectional microphone, combines the sound source positioning space mapping unit, constructs a sound source space distribution model based on the time difference of arrival and the phase difference through a deep learning model, and improves the sound source positioning accuracy. The azimuth angle, pitch angle and distance parameters of the sound source are accurately analyzed, the position of the sound source is mapped to a virtual space coordinate system, a dynamically updated 3D sound source distribution diagram is generated, the position change and intensity distribution of the sound source are reflected in real time, and the positioning precision in a complex acoustic environment is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, specifically to a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction. Background Technology

[0002] With the diversification of modern office and business activities, the forms of meetings are becoming increasingly diverse, including not only traditional face-to-face meetings but also video conferencing, remote meetings, and other forms. However, in some large conference rooms, auditoriums, or halls, due to factors such as large space, hard walls, and poor sound equipment, sound transmission may be distorted and have echoes. These factors can affect the auditory experience of the participants. Especially in complex acoustic environments, ordinary speakers cannot distribute sound evenly, causing participants who are far from the sound source to not be able to hear the content of the speech clearly.

[0003] In multi-source environments, traditional conference sound reinforcement systems struggle to distinguish target speech from interfering sound sources when processing pure audio, especially in multi-person discussion scenarios where erroneous noise reduction is common, leading to decreased speech recognition accuracy. Therefore, the problem this invention aims to solve is how to model the noise reduction process as a Markov decision process, optimize sound source localization accuracy through reinforcement learning, and combine it with an audio engine to more accurately locate and separate different sound sources. To this end, a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction includes a conference management center, which is communicatively connected to the following modules: The multi-sound source perception module is used to capture 360° omnidirectional sound field information in the conference environment, construct a sound source spatial distribution model, map the sound source positions to the virtual space coordinate system, and generate a 3D sound source distribution map. The sound source separation and noise reduction module is used to distinguish target speech from interference sources in the conference environment based on the spatial distribution model of sound sources, and to perform noise reduction processing to achieve high-precision speech enhancement. An adaptive audio engine, combined with a sound source spatial distribution model, uses beamforming and virtual sound source technology to directionally project the separated speech to the target area. The multimodal feedback control module is used to monitor the system's operating status in real time, collect user feedback and system operation data, and further optimize the system's performance.

[0006] A further improvement of the technical solution of the present invention is that the multi-sound source sensing module includes an omnidirectional microphone array unit and a sound source localization spatial mapping unit; The omnidirectional microphone array unit is used to employ a distributed microphone array to cover all directions of the conference space, synchronously collect audio data streams, eliminate the directional limitations of traditional unidirectional microphones, and achieve sound source capture without blind spots. The sound source localization spatial mapping unit constructs a sound source spatial distribution model based on the time difference of arrival and phase difference, combined with a deep learning model, analyzes the azimuth and distance of the sound source, maps the sound source location to a virtual spatial coordinate system, and generates a 3D sound source distribution map.

[0007] A further improvement to the technical solution of the present invention is that the omnidirectional microphone array unit specifically includes: Within the meeting space, multiple microphones are evenly distributed according to the geometric features of the meeting space to form an omnidirectional distributed microphone array. Each microphone is responsible for collecting audio signals in its corresponding direction, ensuring coverage of all directions of the space and achieving omnidirectional audio acquisition. The microphone array is activated, all microphones work synchronously, each microphone independently acquires the original audio signal, multi-channel audio signals are acquired in real time, and the multi-channel audio signals are time-synchronized, the transmission path differences are compensated, and a time-domain aligned audio data stream is generated. The generated audio data stream is preprocessed and integrated, and then synchronously transmitted to the sound source localization spatial mapping unit through the transmission line, ensuring signal integrity and low latency during transmission.

[0008] A further improvement to the technical solution of the present invention is that the sound source localization spatial mapping unit specifically includes: Based on time-domain aligned audio data acquired by a microphone array, the arrival time difference and phase difference between each channel are calculated, and frequency domain features are extracted through short-time Fourier transform to construct a multi-dimensional acoustic feature matrix. The preprocessed arrival time difference, phase difference, and frequency domain features are input into a sound source spatial distribution model constructed by a hybrid 3D convolutional neural network and long short-term memory network. Multimodal information is fused through a spatial attention mechanism to directly regress the azimuth, elevation, and distance parameters of the sound source and perform end-to-end sound source spatial coordinate calculation. The calculated spatial coordinates of the sound sources are mapped to a preset virtual spatial coordinate system. Kalman filtering is then used to smooth the positioning results of consecutive frames in a temporal manner to eliminate transient interference and generate a dynamically updated 3D sound source distribution map, which reflects the changes in the position and intensity distribution of the sound sources in the conference space in real time.

[0009] A further improvement to the technical solution of this invention lies in the following: the construction process of the multi-dimensional acoustic feature matrix is ​​as follows: Based on the time-domain aligned audio data acquired by the microphone array, time-domain alignment calibration is performed to eliminate hardware delay differences. Then, each channel signal is subjected to frame segmentation processing with a frame length of 256 points and a frame shift of 50%. Finally, the time-domain signal is converted into a frequency-domain representation through short-time Fourier transform to obtain a complex spectrum matrix containing amplitude and phase information. The arrival time difference between each channel is calculated using the generalized cross-correlation algorithm (GCC-PHAT). Phase difference information is obtained through frequency domain phase difference unwrapping technology. At the same time, frequency domain features including spectral amplitude, phase and spectral centroid are extracted from the short-time Fourier transform results. The time domain frame index and frequency point dimension are aligned, and the time delay, phase difference and frequency domain features are spliced ​​according to the channel dimension to construct a multi-dimensional acoustic feature matrix containing spatial-temporal-frequency domain information to ensure the spatiotemporal correlation between features. The multidimensional acoustic feature matrix is ​​normalized to eliminate dimensional differences and improve the model convergence speed. Based on the microphone array topology, the multidimensional acoustic feature matrix is ​​spatially reordered to output the standardized multidimensional acoustic feature matrix.

[0010] A further improvement of the technical solution of the present invention is that: the sound source separation and noise reduction module includes a sound source separation unit and an enhanced noise reduction unit; The sound source separation unit is used to separate different sound sources in mixed audio by using the Transformer model and guided by the sound source spatial distribution model. The enhanced noise reduction unit is used to model the noise reduction process as a Markov decision process, using the speech quality score as the reward function, and dynamically adjusts the noise reduction parameters through reinforcement learning (PPO algorithm) to adaptively optimize the noise reduction strategy and avoid speech distortion caused by excessive noise reduction.

[0011] A further improvement to the technical solution of the present invention is that the sound source separation unit specifically includes: By leveraging the prior geometric topology of the microphone array and the spatial distribution model of the sound source, a spatial attention mask is generated. Combined with frequency domain features, it is encoded into a joint spatial-frequency domain feature representation through lightweight CNN, providing structured input for the Transformer model. Based on the multi-head self-attention mechanism, the Transformer model is used to dynamically learn the correlation between different sound sources in the space-frequency domain. The spectral components of each sound source are separated by global context aggregation. Spatial distribution guided loss is introduced to strengthen the correspondence between sound source direction and frequency domain features and improve separation accuracy. The separation features output by the Transformer model are residually connected with the frequency domain features. The time domain signal is reconstructed by inverse short-time Fourier transform. Wiener filtering is then used to further suppress residual interference, resulting in a high-fidelity separated sound source waveform.

[0012] A further improvement to the technical solution of the present invention is that the enhanced noise reduction unit specifically includes: The noise reduction process is modeled as a Markov decision process, with the state space defined as the feature representation of the current audio signal, the action space as the adjustment strategy for the noise reduction parameters, and the reward function as the speech quality score, which is used to evaluate the noise reduction effect. The algorithm employs Proximal Policy Optimization (PPO) from reinforcement learning to dynamically adjust denoising parameters. By sampling audio data under the current policy, calculating reward values, updating the policy network, and optimizing the denoising strategy, the algorithm adaptively adjusts denoising parameters based on the optimized denoising strategy. This allows for real-time processing of audio signals, ensuring speech clarity and naturalness while avoiding excessive denoising that could lead to speech distortion, thus improving overall speech quality.

[0013] A further improvement to the technical solution of the present invention is that the adaptive audio engine specifically includes: Based on the spatial distribution model of sound sources, the location information of each sound source is determined. Using beamforming technology, a directional beam is formed through signal processing of the microphone array to focus on the target sound source, enhance the target speech signal, and suppress interference noise from other directions to ensure the clarity and directionality of the speech. By combining sound source location information and employing virtual sound source technology, the location of the target sound source is simulated in a virtual space. Through audio processing algorithms, the quality of the speech signal is optimized, and the volume and timbre parameters of the speech are adjusted to achieve the best auditory effect in the target area, providing a natural and coherent audio experience. The optimized speech signal is then directed to the target area to ensure that the speech is clearly audible in the designated area.

[0014] A further improvement to the technical solution of the present invention is that the multimodal feedback control module specifically includes: Continuously monitor the system's operating status, covering audio signal processing, sound source localization, and beamforming, acquire real-time data on system performance, and collect user feedback on audio quality, speech clarity, and noise suppression. Through user interface interaction, obtain subjective evaluations of the system's performance. Based on monitoring data and user feedback, the system's operational performance is analyzed, parameter configurations are automatically adjusted, audio processing workflows are optimized, overall system performance is improved, and voice communication quality is ensured.

[0015] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention provides a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction. It covers all directions of the conference space through an omnidirectional microphone array unit, synchronously collects audio data streams, eliminates the limitations of traditional unidirectional microphones, and combines a sound source localization spatial mapping unit. Based on the time difference of arrival and phase difference, it uses a deep learning model to construct a sound source spatial distribution model, accurately analyzes the azimuth angle, pitch angle and distance parameters of the sound source, and then maps the sound source position to a virtual spatial coordinate system to generate a dynamically updated 3D sound source distribution map, which reflects the changes in sound source position and intensity distribution in real time, and significantly improves the localization accuracy in complex acoustic environments.

[0016] 2. This invention provides a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction. It adopts the Transformer model and is guided by the spatial distribution model of sound sources to separate different sound sources in mixed audio. It dynamically learns the correlation between sound sources in the spatial and frequency domains through a multi-head self-attention mechanism. Combined with the spatial distribution guided loss function, it strengthens the correspondence between sound source direction and frequency domain features, thereby improving separation accuracy. The separated frequency domain features are residually connected with the original features, and the time domain signal is reconstructed by inverse short-time Fourier transform. Wiener filtering is used to suppress residual interference and output a high-fidelity speech waveform. This effectively solves the problem of false noise reduction of target speech and interference sound sources in multi-person discussion scenarios, ensuring the accuracy of speech recognition. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram illustrating the workflow of the conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction according to the present invention. Figure 2 This is a data flow diagram of the conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction, as presented in this invention. Detailed Implementation

[0019] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, such as Figure 1, Figure 2 As shown, this invention provides a conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction, including a conference management center. The conference management center has the following communication connections: The multi-sound source perception module is used to capture 360° omnidirectional sound field information in the conference environment and construct a sound source spatial distribution model. It maps the sound source positions to a virtual spatial coordinate system and generates a 3D sound source distribution map. The multi-sound source perception module includes an omnidirectional microphone array unit and a sound source localization spatial mapping unit. The omnidirectional microphone array unit is used to cover all directions of the conference space with a distributed microphone array, synchronously acquiring audio data streams, eliminating the directional limitations of traditional unidirectional microphones, and achieving sound source capture without blind spots. In the conference space, multiple microphones are evenly distributed according to the geometric characteristics of the conference space to form an omnidirectional distributed microphone array. Each microphone is responsible for acquiring audio signals in the corresponding direction, ensuring coverage of all directions of the space and achieving all-round audio acquisition. When the microphone array is activated, all microphones work synchronously, each microphone independently acquires the original audio signal, acquires multi-channel audio signals in real time, and performs time synchronization on the multi-channel audio signals to compensate for transmission path differences, generating a time-domain aligned audio data stream. The generated audio data stream is preprocessed and integrated, and then synchronously transmitted to the sound source localization space mapping unit through the transmission line, ensuring signal integrity and low latency during transmission. The specific functions of the omnidirectional microphone array unit are as follows: Based on the geometric characteristics of the meeting venue, microphones are deployed at a uniform density to form an omnidirectional distributed microphone array. This ensures that each direction is covered by an independent acoustic sensor, and each microphone uses an omnidirectional pickup mode, effectively receiving audio signals from a 360° range on the horizontal plane. This eliminates the directional limitations of traditional unidirectional microphones. Upon startup, all microphones synchronously enter working mode, using high-precision clock synchronization technology for nanosecond-level time alignment to ensure that the audio signals acquired by each channel are consistent in the time domain. The real-time acquired multi-channel audio data includes the location, intensity, and spectral characteristics of the sound source. The omnidirectional microphone array unit calculates the relative time delay of each channel using a time difference of arrival algorithm and employs fractional delay filtering to process the audio signals. The signal undergoes subsampling compensation to generate a time-aligned multi-channel data stream. The audio is further compressed dynamically to suppress burst noise. Simultaneously, bandpass filtering removes low-frequency vibrations (such as air conditioning noise) and ultra-high-frequency interference (such as radiation from electronic devices). The processed data stream is then normalized to unify the gain of each channel to the same order of magnitude, preventing sound source localization errors caused by differences in microphone sensitivity. The pre-processed audio data stream is synchronously transmitted to the sound source localization spatial mapping unit via a low-latency, high-bandwidth transmission line. The transmission protocol employs a deterministic transmission mechanism to ensure that data packets from each channel arrive in a fixed time sequence, avoiding phase distortion caused by network jitter. Forward error correction coding and cyclic redundancy check are embedded in the transmission link to automatically repair transmission errors and ensure signal integrity. The sound source localization spatial mapping unit constructs a sound source spatial distribution model based on time difference of arrival (TDOA) and phase difference, combined with a deep learning model. It analyzes the azimuth and distance of the sound source, maps the sound source position to a virtual spatial coordinate system, and generates a 3D sound source distribution map. Based on time-domain aligned audio data collected by the microphone array, it calculates the TDOA and phase difference between each channel, extracts frequency domain features through short-time Fourier transform, and constructs a multi-dimensional acoustic feature matrix. The preprocessed TDOA, phase difference, and frequency domain features are input into the sound source spatial distribution model constructed by a hybrid 3D convolutional neural network and long short-term memory network. Through spatial attention mechanism, it fuses multimodal information, directly regresses the azimuth, pitch, and distance parameters of the sound source, and performs end-to-end sound source spatial coordinate calculation. The calculated sound source spatial coordinates are mapped to a preset virtual spatial coordinate system. Combined with Kalman filtering, the continuous frame localization results are temporally smoothed to eliminate transient interference and generate a dynamically updated 3D sound source distribution map, which reflects the positional changes and intensity distribution of the sound source in the conference space in real time. Furthermore, the process of constructing the multi-dimensional acoustic feature matrix is ​​as follows: Based on time-domain aligned audio data acquired by a microphone array, time-domain alignment calibration is performed to eliminate hardware latency differences. Each channel signal undergoes framing processing with a frame length of 256 points and a frame shift of 50%. Then, a short-time Fourier transform is used to convert the time-domain signal into a frequency-domain representation, obtaining a complex spectrum matrix containing amplitude and phase information. The generalized cross-correlation algorithm (GCC-PHAT) is used to calculate the arrival time difference between channels. Phase difference information is obtained through frequency-domain phase difference unwrapping technology. Simultaneously, frequency-domain features including spectral amplitude, phase, and spectral centroid are extracted from the short-time Fourier transform results. The system uses a time-domain frame index and frequency dimension to concatenate time delay, phase difference, and frequency domain features along the channel dimension, constructing a multi-dimensional acoustic feature matrix containing spatial-temporal-frequency domain information. This ensures the spatiotemporal correlation between features. The multi-dimensional acoustic feature matrix is ​​then normalized to eliminate dimensional differences and improve model convergence speed. Based on the microphone array topology, the multi-dimensional acoustic feature matrix is ​​spatially reordered, outputting a standardized multi-dimensional acoustic feature matrix with dimensions of number of channels × number of channels × number of feature types × number of time frames. This matrix can be directly input into the sound source spatial distribution model for sound source localization and trajectory analysis. The specific tasks of the sound source localization spatial mapping unit are as follows: Based on the time-domain aligned audio data collected by the microphone array, the arrival time difference and phase difference between each channel are calculated. The arrival time difference is extracted using a generalized cross-correlation algorithm, and the phase difference is calculated using frequency domain phase information from short-time Fourier transform. Both reflect the propagation characteristics of the sound source in space. Simultaneously, the audio data undergoes short-time Fourier transform frame-by-frame processing to extract frequency domain features of spectral amplitude, phase, and spectral centroid, constructing a multi-dimensional acoustic feature matrix containing time delay, phase, and frequency domain information. The preprocessed arrival time difference, phase difference, and frequency domain feature matrix are input into a sound source spatial distribution model constructed using a hybrid 3D convolutional neural network and a long short-term memory network. The 3D convolutional neural network extracts the spatial-temporal correlation of acoustic features through spatiotemporal convolution kernels, capturing local patterns of sound source propagation. The long short-term memory network processes continuous frame features... Temporal modeling is performed to handle the dynamic characteristics of sound source motion, and a spatial attention mechanism is introduced to adaptively fuse multimodal information. The sound source spatial distribution model directly regresses the azimuth, pitch, and distance parameters of the sound source through a fully connected layer, realizing end-to-end sound source spatial coordinate calculation. This avoids the error accumulation of multi-stage processing in traditional methods, significantly improving positioning efficiency and accuracy. The calculated sound source spatial coordinates are mapped to a preset virtual spatial coordinate system, and Kalman filtering is used to temporally smooth the positioning results of continuous frames. Kalman filtering suppresses positioning jitter caused by transient interference (reflected sound, noise abrupt changes, etc.) through iterative updates of the state equation and observation equation, generating a stable spatial trajectory. At the same time, the distribution map is weighted and rendered according to the sound source intensity (spectral energy), reflecting the position changes and intensity distribution of the sound source in the conference space in real time, and then outputting a dynamically updated 3D sound source distribution map in a visual form. The sound source separation and noise reduction module is used to distinguish target speech from interference sources in the conference environment based on the spatial distribution model of sound sources, and to perform noise reduction processing to achieve high-precision speech enhancement. The sound source separation and noise reduction module includes a sound source separation unit and an enhancement and noise reduction unit. The sound source separation unit employs a Transformer model, guided by a sound source spatial distribution model, to separate different sound sources in mixed audio. It utilizes the microphone array's geometric topology and the sound source spatial distribution model as priors to generate a spatial attention mask. Combined with frequency domain features, this mask is encoded into a joint spatial-frequency domain feature representation using a lightweight CNN, providing structured input to the Transformer model. Based on a multi-head self-attention mechanism, the Transformer model dynamically learns the spatial-frequency domain correlations of different sound sources. Global context aggregation separates the spectral components of each sound source, introducing a spatial distribution-guided loss to strengthen the correspondence between sound source direction and frequency domain features, improving separation accuracy. The separation features output by the Transformer model are residually connected with the frequency domain features, and the time-domain signal is reconstructed using an inverse short-time Fourier transform. Wiener filtering further suppresses residual interference, outputting a high-fidelity separated sound source waveform. The specific tasks of the sound source separation unit are as follows: Based on the prior knowledge of the microphone array's geometric topology (element spacing, arrangement shape) and the spatial distribution model of the sound sources (the possible azimuth range of the sound sources), a spatial attention mask is generated. This spatial attention mask calculates the relative direction of each spatial location to the array reference point, and, combined with the beamforming theory of sound source propagation, assigns weights to different spatial regions to highlight the frequency domain features of potential sound source directions. Simultaneously, frequency domain features are extracted from the signals collected by the microphone array and multiplied point-by-point with the spatial attention mask to generate spatially enhanced frequency domain features. A lightweight CNN (1D convolution) is used to encode the joint spatial-frequency features, capturing the correlation between frequency bands through local receptive fields, and outputting a structured feature representation as input to the Transformer model. Based on a multi-head self-attention mechanism, the Transformer model dynamically learns the spatial-frequency correlation of different sound sources, decomposing the input features into... Multiple subspaces (attention heads) are used, each independently calculating global attention weights for spatial location and frequency points to capture the cross-dimensional dependency between sound source direction and spectral components. Through global context aggregation, the spectral components of each sound source are separated to generate independent frequency domain representations. Simultaneously, a spatial distribution-guided loss function is introduced, which is optimized based on the difference between the actual sound source location and the spatial attention distribution predicted by the model. This forces the model to follow spatial prior constraints when separating the spectrum, improving separation accuracy. The separated frequency domain features output by the Transformer model are residually connected with the original frequency domain features to retain detailed information in the original signal and accelerate model convergence. The frequency domain representation is reconstructed into a time domain signal through inverse short-time Fourier transform. The reconstructed signal is further post-processed by Wiener filtering: the optimal filter coefficients are estimated based on the signal-to-noise ratio of the separated spectrum to suppress the spectral components of non-target sound sources while preserving the phase continuity of the target signal, outputting a high-fidelity separated sound source waveform. The enhanced noise reduction unit models the noise reduction process as a Markov decision process, using speech quality scores as the reward function. It dynamically adjusts noise reduction parameters through reinforcement learning (PPO algorithm) to adaptively optimize the noise reduction strategy, avoiding speech distortion caused by excessive noise reduction. The noise reduction process is modeled as a Markov decision process, defining the state space as the feature representation of the current audio signal, the action space as the adjustment strategy for noise reduction parameters, and the reward function as the speech quality score to evaluate the noise reduction effect. It employs the Proximal Policy Optimization (PPO) algorithm in reinforcement learning to dynamically adjust noise reduction parameters. By sampling audio data under the current strategy, calculating the reward value, updating the policy network, and optimizing the noise reduction strategy, it adaptively adjusts the noise reduction parameters based on the optimized strategy, processing the audio signal in real time. This ensures speech clarity and naturalness while avoiding speech distortion caused by excessive noise reduction, thus improving overall speech quality. The specific work of the enhanced noise reduction unit is as follows: The dynamic parameter adjustment problem is modeled as a Markov decision process. The state space consists of the time-frequency characteristics of the current audio signal and historical noise reduction parameters to capture the time-varying characteristics of the signal and the continuity of parameter adjustment. The action space is defined as the incremental adjustment strategy of the noise reduction parameters, including the step change range of key parameters such as noise suppression intensity, frequency band filtering threshold, and phase compensation coefficient, ensuring the continuity and differentiability of the actions. Through the Markov decision process framework, the noise reduction process is transformed into a sequential decision problem. At each time step, the optimal action is selected based on the current state, balancing noise suppression and speech fidelity. State transitions are driven by the dynamic characteristics of the audio signal, and the execution of actions affects the signal state at the next moment, forming a closed-loop control structure. The PPO algorithm is used to dynamically optimize the noise reduction strategy. The policy network and value network of the PPO algorithm are analyzed. The policy network takes state features as input and outputs the probability distribution of actions, while the value network evaluates the long-term reward of the current state-action pair. In each training iteration, the algorithm interacts with the environment (i.e., processes real-time audio). The algorithm collects trajectory data, calculates the reward value (based on speech quality score) and the advantage function estimate for each action, and limits the policy update range by minimizing the pruned surrogate objective function to avoid performance crashes caused by policy abrupt changes. Simultaneously, it utilizes generalized advantage estimation to balance bias and variance, improving the accuracy of gradient estimation and generating dynamic denoising strategies adapted to different noise scenarios. Based on the current state characteristics, it generates denoising parameter adjustment actions and uses smoothing filtering to avoid speech distortion caused by parameter abrupt changes. The reward function is designed as a linear combination of speech quality scores, covering signal-to-noise ratio improvement and speech distortion indicators, ensuring comprehensive optimization of denoising effect and speech naturalness. The strategy is continuously updated through an online learning mechanism: if the current action leads to a decrease in reward, the policy network reduces the sampling probability of that action; conversely, it strengthens the exploration of advantageous actions and introduces an action constraint module to limit the parameter adjustment range, preventing excessive denoising from damaging the speech harmonic structure. Finally, it achieves adaptive parameter calibration in dynamic noise environments, balancing noise suppression and speech fidelity, and improving the quality of speech communication in complex scenarios. An adaptive audio engine, combined with a sound source spatial distribution model, uses beamforming and virtual sound source technology to directionally project the separated speech to the target area. The multimodal feedback control module is used to monitor the system's operating status in real time, collect user feedback and system operation data, and further optimize the system's performance.

[0021] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: preferably, the adaptive audio engine specifically includes: Based on the spatial distribution model of sound sources, the location information of each sound source is determined. Using beamforming technology, a directional beam is formed through signal processing of the microphone array to focus on the target sound source, enhancing the target speech signal while suppressing interference noise from other directions, ensuring the clarity and directionality of the speech. Combining the sound source location information, virtual sound source technology is used to simulate the location of the target sound source in virtual space. Through audio processing algorithms, the quality of the speech signal is optimized, and the volume and timbre parameters of the speech are adjusted to achieve the best auditory effect in the target area, providing a natural and coherent audio experience. The optimized speech signal is then directed to the target area to ensure that the speech is clearly audible in the specified area. The adaptive audio engine works as follows: Based on a spatial distribution model of sound sources, it determines the location information of each sound source. Using beamforming technology and leveraging the signal processing capabilities of the microphone array, it constructs a directional beam that focuses on the target sound source, effectively enhancing the target speech signal while significantly suppressing interference noise from other directions, ensuring the clarity and directionality of the speech. Specifically, signal processing algorithms are used to weight and delay the audio signals collected by the microphone array, ensuring the beam accurately aligns with the target sound source, achieving efficient extraction and enhancement of the target speech. Combining the sound source location information, the adaptive audio engine employs virtual sound source technology to simulate the location of the target sound source in virtual space. Through audio processing algorithms, it optimizes the speech signal, adjusting the volume and timbre of the speech. The parameters are optimized to achieve the best auditory effect for the speech signal within the target area, providing a natural and coherent audio experience. The audio processing algorithm dynamically adjusts the speech signal based on the acoustic environment of the target area and the needs of the listener, ensuring that the volume of the speech is moderate and the timbre is natural, while reducing the impact of echo and noise, and improving the intelligibility and comfort of the speech. The adaptive audio engine projects the optimized speech signal directionally to the target area, ensuring that the speech is clearly audible in the specified area. In particular, through sound field control technology, combined with the sound source location information and the acoustic characteristics of the target area, the speech signal is spatially located and projected. At the same time, the acoustic environment of the target area is monitored in real time, and the projection strategy of the speech signal is automatically adjusted according to the environmental changes to ensure that the speech signal always maintains the best auditory effect within the target area. The multimodal feedback control module specifically includes: Continuously monitor the system's operating status, covering audio signal processing, sound source localization, and beamforming, acquire real-time data on system performance, and collect user feedback on audio quality, speech clarity, and noise suppression. Through user interface interaction, obtain subjective evaluations of the system's performance. Based on monitoring data and user feedback, analyze the system's operating effect, automatically adjust parameter configurations, optimize audio processing workflows, improve overall system performance, and ensure voice communication quality. The multimodal feedback control module's specific functions are as follows: It continuously monitors the system's operating status, covering key aspects of audio signal processing, sound source localization, and beamforming. It acquires real-time data on system performance, including the time-frequency characteristics of the audio signal, the accuracy of sound source localization, and the effectiveness of beamforming. Simultaneously, it collects user feedback on audio quality, speech clarity, and noise suppression, obtaining subjective evaluations of the system's performance through an interactive interface. Based on the monitoring data and user feedback, the multimodal feedback control module conducts in-depth analysis of the system's operating performance. It evaluates the performance of the current audio processing flow using data analysis techniques, identifies potential optimization points, and automatically adjusts system parameter configurations based on the analysis results. This includes the gain of the audio signal processing, the parameters of the sound source localization algorithm, and the direction of beamforming, thereby optimizing the audio processing flow, improving overall system performance, and ensuring optimal voice communication quality.

[0022] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A conference sound reinforcement system based on AI intelligent algorithms and 360° omnidirectional noise reduction, including a conference management center, characterized in that: The conference management center communication connection includes the following modules, among which: The multi-sound source perception module is used to capture 360° omnidirectional sound field information in the conference environment, construct a sound source spatial distribution model, map the sound source positions to the virtual space coordinate system, and generate a 3D sound source distribution map. The sound source separation and noise reduction module is used to distinguish target speech from interference sources in the conference environment and perform noise reduction processing, guided by the spatial distribution model of sound sources. An adaptive audio engine, combined with a sound source spatial distribution model, uses beamforming and virtual sound source technology to directionally project the separated speech to the target area. The multimodal feedback control module is used to monitor the system's operating status in real time, collect user feedback and system operation data, and further optimize the system's performance.

2. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 1, characterized in that: The multi-sound source sensing module includes an omnidirectional microphone array unit and a sound source localization spatial mapping unit. The omnidirectional microphone array unit is used to employ a distributed microphone array to cover all directions of the conference space and synchronously collect audio data streams. The sound source localization spatial mapping unit constructs a sound source spatial distribution model based on the time difference of arrival and phase difference, combined with a deep learning model, analyzes the azimuth and distance of the sound source, maps the sound source location to a virtual spatial coordinate system, and generates a 3D sound source distribution map.

3. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 2, characterized in that: The omnidirectional microphone array unit specifically includes: Within the meeting space, multiple microphones are evenly distributed according to the geometric features of the meeting space, forming an omnidirectional distributed microphone array; The microphone array is activated, all microphones work synchronously, each microphone independently acquires the original audio signal, multi-channel audio signals are acquired in real time, and the multi-channel audio signals are time-synchronized, the transmission path differences are compensated, and a time-domain aligned audio data stream is generated. The generated audio data stream is preprocessed and integrated, and then synchronously transmitted to the sound source localization spatial mapping unit through the transmission line.

4. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 2, characterized in that: The sound source localization spatial mapping unit specifically includes: Based on time-domain aligned audio data acquired by a microphone array, the arrival time difference and phase difference between each channel are calculated, and frequency domain features are extracted through short-time Fourier transform to construct a multi-dimensional acoustic feature matrix. The preprocessed arrival time difference, phase difference, and frequency domain features are input into a sound source spatial distribution model constructed by a hybrid 3D convolutional neural network and long short-term memory network. Multimodal information is fused through a spatial attention mechanism to directly regress the azimuth, elevation, and distance parameters of the sound source and perform end-to-end sound source spatial coordinate calculation. The calculated spatial coordinates of the sound sources are mapped to a preset virtual spatial coordinate system, and Kalman filtering is used to perform temporal smoothing on the localization results of consecutive frames to generate a dynamically updated 3D sound source distribution map.

5. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 4, characterized in that: The process of constructing the multi-dimensional acoustic feature matrix is ​​as follows: Based on the time-domain aligned audio data acquired by the microphone array, time-domain alignment calibration is performed, and each channel signal is subjected to frame segmentation processing with a frame length of 256 points and a frame shift of 50%. Then, the time-domain signal is converted into a frequency-domain representation through short-time Fourier transform to obtain a complex spectrum matrix containing amplitude and phase information. The arrival time difference between each channel is calculated using a generalized cross-correlation algorithm. Phase difference information is obtained through frequency domain phase difference unwrapping technology. At the same time, frequency domain features including spectral amplitude, phase and spectral centroid are extracted from the short-time Fourier transform results. The time domain frame index and frequency point dimension are aligned, and the time delay, phase difference and frequency domain features are spliced ​​according to the channel dimension to construct a multi-dimensional acoustic feature matrix containing spatial-temporal-frequency domain information. The multi-dimensional acoustic feature matrix is ​​normalized and spatially reordered according to the microphone array topology to output the standardized multi-dimensional acoustic feature matrix.

6. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 2, characterized in that: The sound source separation and noise reduction module includes a sound source separation unit and an enhanced noise reduction unit; The sound source separation unit is used to separate different sound sources in mixed audio by using the Transformer model and guided by the sound source spatial distribution model. The enhanced noise reduction unit is used to model the noise reduction process as a Markov decision process, using speech quality score as the reward function, and dynamically adjust the noise reduction parameters through reinforcement learning to adaptively optimize the noise reduction strategy.

7. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 6, characterized in that: The sound source separation unit specifically includes: By leveraging the prior geometric topology of the microphone array and the spatial distribution model of the sound source, a spatial attention mask is generated. Combined with frequency domain features, it is encoded into a joint spatial-frequency domain feature representation through a lightweight CNN. Based on the multi-head self-attention mechanism, the Transformer model is used to dynamically learn the correlation between different sound sources in the space-frequency domain. The spectral components of each sound source are separated by global context aggregation, and spatial distribution guided loss is introduced to strengthen the correspondence between sound source direction and frequency domain features. The separation features output by the Transformer model are residually connected with the frequency domain features. The time domain signal is reconstructed by inverse short-time Fourier transform. Wiener filtering is then used to further suppress residual interference, resulting in a high-fidelity separated sound source waveform.

8. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 6, characterized in that: The enhanced noise reduction unit specifically includes: The noise reduction process is modeled as a Markov decision process, with the state space defined as the feature representation of the current audio signal, the action space as the adjustment strategy for the noise reduction parameters, and the reward function as the speech quality score. The algorithm employs a proximal policy optimization algorithm in reinforcement learning to dynamically adjust the noise reduction parameters. By sampling audio data under the current policy, calculating the reward value, updating the policy network, and optimizing the noise reduction policy, the algorithm adaptively adjusts the noise reduction parameters based on the optimized noise reduction policy to process the audio signal in real time.

9. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 6, characterized in that: The adaptive audio engine specifically includes: Based on the spatial distribution model of sound sources, the location information of each sound source is determined. Then, using beamforming technology and signal processing of a microphone array, a directional beam is formed and focused on the target sound source. By combining sound source location information and using virtual sound source technology, the location of the target sound source is simulated in virtual space. Then, through audio processing algorithms, the quality of the speech signal is optimized, and the volume and timbre parameters of the speech are adjusted to achieve the best auditory effect in the target area. Finally, the optimized speech signal is directed to the target area.

10. The conference sound reinforcement system based on AI intelligent algorithm and 360° omnidirectional noise reduction according to claim 9, characterized in that: The multimodal feedback control module specifically includes: Continuously monitor the system's operating status, covering audio signal processing, sound source localization, and beamforming, acquire real-time data on system performance, and collect user feedback on audio quality, speech clarity, and noise suppression. Through user interface interaction, obtain subjective evaluations of the system's performance. Based on monitoring data and user feedback, the system's performance is analyzed, parameter configurations are automatically adjusted, and the audio processing workflow is optimized.

Images