Digital conference power adaptive adjustment method based on AI noise identification

Abstract

The application relates to the technical field of digital conference systems, and discloses a digital conference power self-adaptive adjustment method based on AI noise recognition. The method collects conference environment noise data through a distributed microphone array, carries out voiceprint feature extraction and spectrum decomposition, generates a multi-dimensional noise feature spectrum, constructs a convolution-circulation hybrid recognition model to generate a noise interference level index, generates a signal-to-noise ratio degradation trend diagram based on a dynamic threshold model, constructs a power adjustment decision matrix, carries out online incremental learning on the model after multi-target parameter optimization, divides conference terminal power levels according to the output of the updated model, generates a dynamic gain threshold and constraints, obtains a partition adjustment strategy set, adapts to the parameters of an audio processing unit, generates a real-time power adjustment signal and carries out logical verification, and finally obtains a final power adjustment scheme. The method can adaptively adjust conference power, improve conference voice quality and enhance the environmental adaptability of the system.

Description

Technical Field

[0001] This invention relates to the field of digital conferencing system technology, specifically to a digital conferencing power adaptive adjustment method based on AI noise recognition. Background Technology

[0002] In modern digital conferencing systems, noise interference in the meeting environment has always been a key issue affecting meeting quality. With the widespread use of remote conferencing, video conferencing, and other similar formats, meeting scenarios have become increasingly complex, and noise sources have become more diverse, including steady-state environmental noise, localized interference noise, and transient burst noise. These noises can degrade the signal-to-noise ratio of audio signals, severely impacting the listening experience of meeting participants and the efficiency of information transmission.

[0003] Traditional methods for adjusting conference power often rely on fixed thresholds or simple manual adjustments, making them ill-suited to complex and ever-changing conference environments. For example, when faced with sudden noise interference, traditional methods cannot respond promptly, resulting in untimely power adjustments and further exacerbating audio signal distortion. Furthermore, traditional methods typically fail to consider the differences in acoustic characteristics across various conference scenarios, such as the size, shape, and materials of the venue, making the power adjustment solutions lack environmental adaptability and unable to achieve satisfactory results in diverse situations.

[0004] Furthermore, existing noise identification technologies often fail to fully extract key noise features, such as background noise fundamental frequency, sudden interference pulses, and speech aliasing, when processing multi-dimensional noise characteristics. This leads to inaccurate noise interference level assessments, which in turn affects the accuracy of power regulation. Moreover, traditional power regulation decision-making processes lack dynamic optimization mechanisms, making it impossible to adjust and optimize the regulation strategy in a timely manner based on real-time noise changes, resulting in poor system adaptability.

[0005] In terms of power control for conference terminals, traditional methods typically employ a uniform power adjustment strategy, failing to consider the differences in location and function of different terminals within a conference scenario. This leads to unreasonable power allocation, affecting not only conference quality but also potential energy waste. Furthermore, traditional methods lack logical verification and security assessment of adjustment signals during power adjustment, making them prone to issues such as adjustment anomalies, impacting system stability and reliability.

[0006] With the rapid development of artificial intelligence (AI) technology, applying AI to noise identification and power regulation in digital conferencing systems has become an important research direction. However, current research still has many shortcomings, such as the difficulty in balancing real-time performance and accuracy of AI models, insufficient online update capabilities, and inability to adapt to constantly changing noise environments. Therefore, there is an urgent need for an AI-based adaptive power regulation method for digital conferencing systems to improve their noise immunity and the intelligence level of power regulation, thus meeting the demands of modern conferences for high-quality voice communication. Summary of the Invention

[0007] The purpose of this invention is to provide a digital conference power adaptive adjustment method based on AI noise recognition, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a digital conference power adaptive adjustment method based on AI noise recognition, the method comprising:

[0009] The conference environment noise data is collected in real time by a distributed microphone array, and the noise data is subjected to voiceprint feature extraction and spectrum decomposition to generate a multi-dimensional noise feature spectrum.

[0010] Convolutional cyclic hybrid recognition model is constructed by inputting background noise fundamental frequency, sudden interference pulses and speech aliasing data from the multi-dimensional noise feature spectrum, and generating a noise interference level index.

[0011] The noise interference level index is analyzed in real time based on the dynamic threshold model, a signal-to-noise ratio degradation trend map is generated, and a power adjustment decision matrix is ​​constructed based on the signal-to-noise ratio degradation trend map.

[0012] The power adjustment decision matrix is ​​optimized for multiple objectives to generate a model update instruction, and the convolutional recurrent hybrid recognition model is subjected to online incremental learning based on the model update instruction.

[0013] Based on the updated model output, the power levels of the conference terminals are divided, dynamic gain thresholds are generated, and environmental adaptability constraints are applied to the dynamic gain thresholds to obtain a set of partition adjustment strategies.

[0014] The audio processing unit is parameter-adapted according to the partition adjustment strategy set to generate a real-time power adjustment signal, and the real-time power adjustment signal is logically verified to generate the final power adjustment scheme.

[0015] Preferably, the real-time acquisition of conference environment noise data through a distributed microphone array includes:

[0016] Steady-state environmental noise data was collected using a top-mounted microphone in the venue, local interference noise data was collected using a desktop directional microphone, and transient burst noise data was obtained using a mobile terminal microphone.

[0017] The steady-state environmental noise data, local interference noise data, and transient burst noise data are processed by timestamp alignment and frequency band normalization to generate a synchronous multi-channel audio stream.

[0018] The synchronous multi-channel audio stream is subjected to background noise cancellation and echo retention suppression to obtain purified multi-source noise data.

[0019] Preferably, the construction of the convolutional recurrent hybrid recognition model includes:

[0020] Speech proficiency constraints and energy conservation equations are defined as boundary conditions for the model, and a joint processor containing temporal convolutional modules and gated recurrent units is constructed.

[0021] Harmonic decomposition is performed on the fundamental frequency of the background noise to generate noise source features; envelope detection is performed on the sudden interference pulse to generate impact duration features; and spectral entropy is calculated on the speech aliasing data to generate spectral pollution features.

[0022] The noise source characteristics, impact duration characteristics, and spectral pollution characteristics are input into the joint processor for feature fusion, and a noise interference level index is output.

[0023] Preferably, the real-time analysis of the noise interference level index based on the dynamic threshold model includes:

[0024] An environmental fitness comparison table was constructed based on historical conference noise samples, and the noise interference level index was sampled using a rolling time window to generate an index fluctuation sequence.

[0025] The index fluctuation sequence is fitted with a trend using a double exponential smoothing algorithm to generate dynamic threshold parameters. The cumulative distribution of the dynamic threshold parameters is then calculated based on the environmental fitness reference table to obtain a signal-to-noise ratio degradation trend map.

[0026] Preferably, the multi-objective parameter optimization of the power regulation decision matrix includes:

[0027] The baseline gain value is calculated based on the speaker response curve and room acoustic characteristics, and the slope of the signal-to-noise ratio degradation trend graph is detected to generate the adjustment demand increment.

[0028] The baseline gain value and the adjustment demand increment are Pareto optimized using a non-dominated sorting genetic algorithm to generate an updated power adjustment decision matrix, and the model update instruction is determined based on the matrix singular value decomposition results.

[0029] Preferably, the step of classifying the power level of the conference terminal based on the updated model output includes:

[0030] Regional clustering analysis is performed on the noise interference level index to generate a set of high interference partitions, and frequency band correlation calibration is performed on the dynamic gain threshold to generate environmental adaptability constraints.

[0031] Based on the environmental adaptability constraints, a topological mapping is performed on the set of high-interference partitions to generate a set of partition adjustment strategies.

[0032] Preferably, the generation of the dynamic gain threshold includes:

[0033] Gaussian mixture modeling is performed on the noise interference level index to generate a noise tolerance range, and differential calculation is performed on the multi-dimensional noise feature spectrum to generate abnormal voiceprint features.

[0034] The abnormal voiceprint features are marked with abrupt change regions using an adaptive threshold segmentation algorithm to generate an initial gain threshold. The initial gain threshold is then corrected for sound field uniformity to obtain a dynamic gain threshold.

[0035] Preferably, generating the real-time power adjustment signal includes:

[0036] A priority weight table is constructed based on the terminal location distribution and acoustic coverage requirements, and energy diffusion analysis is performed on the partitioned adjustment strategy set to generate adjustment intensity coefficients.

[0037] The adjustment intensity coefficient is discretized and encoded using a fuzzy state transition machine to generate multi-level adjustment signals. Time delay compensation processing is then applied to the multi-level adjustment signals to obtain real-time power adjustment signals.

[0038] Preferably, the generation of the final power regulation scheme includes:

[0039] The real-time power adjustment signal is matched for device compatibility to generate a basic adjustment strategy, and the basic adjustment strategy is verified for power consumption limitations to generate a feasible adjustment instruction set.

[0040] The feasible adjustment instruction set is verified by sound field reconstruction through digital twin simulation to generate the final power adjustment scheme.

[0041] Preferably, the logical verification of the real-time power adjustment signal includes:

[0042] The real-time power adjustment signal is cross-compared through a redundant verification channel to generate a verification result vector, and the consistency of the verification result vector is evaluated to generate a verification pass flag.

[0043] If the verification pass flag is true, the final power adjustment scheme is executed; if it is false, a safety rollback mechanism is triggered and a backup adjustment scheme is generated.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] The proposed digital conference power adaptive adjustment method based on AI noise recognition uses a distributed microphone array to collect steady-state environmental noise, local interference noise, and transient burst noise data in the conference environment in real time. The data is then processed by timestamp alignment, frequency band normalization, background noise elimination, and echo retention suppression. This method can accurately obtain purified multi-source noise data, providing a reliable data foundation for subsequent noise recognition and power adjustment.

[0046] The constructed convolutional-recurrent hybrid recognition model, which combines a temporal convolution module and a gated recurrent unit, performs harmonic decomposition on the fundamental frequency of background noise, envelope detection on sudden interference pulses, and spectral entropy calculation on speech aliasing data. It can fully extract and fuse key noise features, accurately generate noise interference level indicators, and improve the accuracy and reliability of noise recognition.

[0047] Based on the dynamic threshold model, noise interference level indicators are analyzed in real time. The trend is fitted using the double exponential smoothing algorithm. Combined with the environmental fitness comparison table constructed from historical conference noise samples, an accurate signal-to-noise ratio degradation trend map can be generated, providing a scientific basis for power adjustment decisions.

[0048] By using a non-dominated sorting genetic algorithm to optimize the power adjustment decision matrix through multi-objective parameters, and combining the loudspeaker response curve, room acoustic characteristics, and adjustment demand increments, the optimal power adjustment decision matrix can be generated. Furthermore, the model update instructions are determined through matrix singular value decomposition, enabling online incremental learning of the model and improving the system's adaptability and adjustment accuracy.

[0049] By performing regional clustering analysis and topological mapping on noise interference level indicators, combined with frequency band correlation calibration and environmental adaptability constraints, a set of zonal adjustment strategies can be generated to achieve precise hierarchical classification of conference terminal power, thereby improving the rationality of power allocation and environmental adaptability.

[0050] By generating a dynamic gain threshold through Gaussian mixture modeling, adaptive threshold segmentation algorithm, and sound field uniformity correction, the gain can be dynamically adjusted according to the real-time noise situation, thereby improving the system's noise immunity and speech intelligibility.

[0051] Based on the terminal location distribution and acoustic coverage requirements, a priority weight table is constructed. Combined with energy diffusion analysis and fuzzy state transition machine, a real-time power adjustment signal is generated and time delay compensation processing is performed. This enables precise parameter adaptation of the audio processing unit, improving the real-time performance and accuracy of power adjustment.

[0052] By performing device compatibility matching, power consumption limit verification, and digital twin simulation verification on real-time power regulation signals, a reliable final power regulation scheme can be generated. At the same time, logic verification is performed through redundant verification channels to ensure the safety and reliability of the regulation scheme and improve the stability and reliability of the system.

[0053] Furthermore, this method, through multi-objective parameter optimization and online incremental learning, can continuously optimize the model and adjustment strategy, improve the system's adaptability and performance, reduce the need for manual intervention, and enhance the intelligence level of the conference system. Simultaneously, precise power adjustment can rationally allocate power while ensuring conference quality, reducing energy consumption and demonstrating excellent energy-saving and environmental protection effects. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based noise recognition-based digital conference power adaptive adjustment method described in this invention.

[0055] Figure 2 Design diagram for noise data acquisition and processing;

[0056] Figure 3 Design diagram for dynamic threshold analysis;

[0057] Figure 4 Design diagram for generating real-time power regulation signals. Detailed Implementation

[0058] 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, and 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.

[0059] Please see Figures 1-4 The present invention relates to a digital conference power adaptive adjustment method based on AI noise recognition, the specific implementation steps of which are as follows:

[0060] This method uses a distributed microphone array to collect real-time ambient noise data from a meeting environment. It then extracts voiceprint features and decomposes the noise data to generate a multi-dimensional noise feature spectrum. Specifically, it collects steady-state ambient noise data, such as air conditioning and ventilation system noise, using overhead microphones; it collects local interference noise data, such as keyboard clicks and paper turning sounds, using desktop directional microphones; and it acquires transient burst noise data, such as phone ringtones and sudden coughs, using mobile terminal microphones. Next, it performs timestamp alignment and frequency band normalization on the steady-state ambient noise data, local interference noise data, and transient burst noise data to ensure comparability of noise data from different sources in both time and frequency dimensions, generating a synchronous multi-channel audio stream. Then, it performs background noise cancellation and echo suppression on the synchronous multi-channel audio stream, employing advanced signal processing algorithms to remove inherent noise and echoes from the environment, resulting in purified multi-source noise data.

[0061] A convolutional-recurrent hybrid recognition model is constructed, taking into account background noise fundamental frequency, sudden interference pulses, and speech aliasing data from a multi-dimensional noise feature spectrum, to generate a noise interference level index. First, speech clarity constraints and the energy conservation equation are defined as boundary conditions to ensure the model's output meets the needs of real-world meeting scenarios. Then, a coprocessor containing a temporal convolution module and a gated recurrent unit is constructed. The temporal convolution module extracts local noise features, while the gated recurrent unit processes the temporal features of the noise. Harmonic decomposition is performed on the background noise fundamental frequency to generate noise source features. By analyzing the harmonic components of the fundamental frequency, the source and characteristics of the noise are determined. Envelope detection is performed on sudden interference pulses to generate impulse duration features, describing the intensity and duration of the interference pulses. Spectral entropy is calculated on the speech aliasing data to generate spectral pollution features, reflecting the degree of noise contamination of the speech signal. Finally, the noise source features, impulse duration features, and spectral pollution features are input into the coprocessor for feature fusion. Through model learning and calculation, a noise interference level index is output, which accurately reflects the degree of noise interference in the meeting environment.

[0062] A dynamic threshold model is used to analyze noise interference level indicators in real time, generating a signal-to-noise ratio (SNR) degradation trend map. A power adjustment decision matrix is ​​then constructed based on this trend map. An environmental fitness reference table is built using historical conference noise samples, showing the adaptation under different noise environments. A rolling time window sampling method is applied to the noise interference level indicators to generate an indicator fluctuation sequence. Changes in the noise indicators are captured in real time using a sliding time window. A double exponential smoothing algorithm is used to fit the trend of the indicator fluctuation sequence, generating dynamic threshold parameters. This algorithm effectively eliminates the randomness of noise fluctuations and extracts trend features. The cumulative distribution of the dynamic threshold parameters is calculated based on the environmental fitness reference table, resulting in a SNR degradation trend map. This map visually displays the SNR change trend over time, providing a basis for subsequent power adjustment.

[0063] Multi-objective parameter optimization is performed on the power adjustment decision matrix to generate model update instructions, and online incremental learning is performed on the convolutional recurrent hybrid recognition model based on these instructions. A baseline gain value is calculated based on the speaker response curve and room acoustic characteristics to ensure that power adjustment meets the basic requirements of the equipment and environment. Slope detection is performed on the signal-to-noise ratio degradation trend graph to generate adjustment demand increments. By analyzing the slope of the trend graph, the urgency and magnitude of power adjustment are determined. Pareto optimization is performed on the baseline gain value and adjustment demand increments using a non-dominated sorting genetic algorithm. This algorithm can find the optimal balance point among multiple objectives, generating an updated power adjustment decision matrix. Model update instructions are determined based on the matrix singular value decomposition results. Through matrix decomposition and analysis, the parameters and directions that need to be updated by the model are determined, enabling online incremental learning of the model and allowing it to continuously adapt to new noise environments.

[0064] Based on the updated model output, the power levels of the conference terminals are classified, dynamic gain thresholds are generated, and environmental adaptability constraints are applied to these dynamic gain thresholds to obtain a set of zone adjustment strategies. Regional clustering analysis is performed on the noise interference level indicators, and the conference venue is divided into different regions using a clustering algorithm to generate a set of high-interference zones. Frequency band correlation calibration is performed on the dynamic gain thresholds, considering the correlation between different frequency bands to generate environmental adaptability constraints. Based on these constraints, topological mapping is performed on the high-interference zone set, mapping the high-interference areas onto the topology of the conference venue to generate a set of zone adjustment strategies. This set of strategies specifies corresponding power adjustment strategies for different high-interference areas.

[0065] When generating the dynamic gain threshold, Gaussian mixture modeling is applied to the noise interference level index to generate a noise tolerance range. The Gaussian mixture model describes the distribution of the noise index, determining the noise range that the system can tolerate. Differential calculation is performed on the multi-dimensional noise feature spectrum to generate abnormal voiceprint features, and abrupt changes in noise features are detected through differential operations. An adaptive threshold segmentation algorithm is used to mark abrupt change regions of abnormal voiceprint features, generating an initial gain threshold. This algorithm can automatically adjust the threshold according to changes in noise features. Sound field uniformity correction is applied to the initial gain threshold, considering the uniformity of the sound field in the conference venue, to obtain the dynamic gain threshold, ensuring consistent sound effects across all areas.

[0066] When generating real-time power regulation signals, a priority weight table is constructed based on the terminal location distribution and acoustic coverage requirements, assigning different priority weights to terminal devices in different locations. Energy diffusion analysis is performed on the zonal regulation strategy set to generate regulation intensity coefficients, analyzing the energy diffusion of power regulation in different areas to determine the regulation intensity. The regulation intensity coefficients are discretized and encoded using a fuzzy state transition machine to generate multi-level regulation signals, transforming continuous regulation intensity into discrete multi-level signals. Time delay compensation processing is applied to the multi-level regulation signals, considering the time delay during signal transmission to obtain real-time power regulation signals, ensuring the real-time performance and accuracy of power regulation.

[0067] When generating the final power regulation scheme, device compatibility matching is performed on the real-time power regulation signal to generate a basic regulation strategy, ensuring compatibility between the regulation signal and the device. Power consumption limitation verification is performed on the basic regulation strategy to generate a feasible regulation instruction set, controlling device power consumption while meeting power regulation requirements. The feasible regulation instruction set is verified through sound field reconstruction using digital twin simulation. A virtual model of the conference venue is constructed using digital twin technology, and the regulation instruction set is simulated and verified to generate the final power regulation scheme, ensuring the effectiveness and reliability of the scheme.

[0068] When performing logical verification on the real-time power regulation signal, a redundant verification channel is used to cross-compare the signal, generating a verification result vector. Multiple verification channels are utilized to improve verification accuracy. A consistency evaluation is performed on the verification result vector, generating a verification pass flag. If the verification pass flag is true, the final power regulation scheme is executed; if false, a safety fallback mechanism is triggered, and a backup regulation scheme is generated to ensure system safety and stability.

[0069] Example 1:

[0070] When collecting environmental noise data for the conference, steady-state environmental noise data is collected using overhead microphones. These microphones are typically installed in suitable locations on the conference room ceiling, enabling comprehensive collection of continuous noise throughout the entire conference room environment. For example, the continuous low-frequency noise generated by the conference room air conditioning is stable and persistent, constituting a significant component of steady-state environmental noise; the airflow noise from the ventilation system is also considered steady-state environmental noise and can cause continuous background interference to conference speech. Localized interference noise data is collected using desktop directional microphones, which are generally placed at various locations on the conference table to collect noise data from localized areas near the participants. High-frequency sounds from participants typing on keyboards are clearly localized and sudden, interfering with nearby speech collection; the rustling sound of turning over paper documents is also considered localized interference noise, affecting the clarity of speech in localized areas. Transient burst noise data is acquired using mobile terminal microphones, including microphones from participants' mobile phones, tablets, and other devices, which can capture sudden, brief noises in the conference environment. For example, a sudden ringing of a cell phone is random and unexpected, and its intensity is relatively high, which can seriously interfere with the meeting; another example is an unintentional cough by a participant, which is also a transient noise that can interfere with the voice signal for a short period of time.

[0071] After acquiring these different types and sources of noise data, it is necessary to perform timestamp alignment processing on the steady-state environmental noise data, local interference noise data, and transient burst noise data. Since there may be slight differences in the sampling times of different microphones, timestamp alignment processing involves adding a precise timestamp to each noise data sample and employing a time synchronization algorithm to ensure that the data collected by different microphones are strictly synchronized on the time axis. This ensures that the data samples are consistent in time when subsequent multi-channel noise data is analyzed and processed, avoiding deviations in analysis results due to time differences. Simultaneously, frequency band normalization processing is performed. Different types of noise may be distributed in different frequency ranges; frequency band normalization unifies the frequency range of all noise data to the same standard range. Through signal processing techniques such as filtering and resampling, noise data in different frequency bands are made comparable in the frequency dimension, facilitating subsequent feature extraction and model processing to generate a synchronized multi-channel audio stream.

[0072] Next, background noise cancellation is performed on the synchronous multi-channel audio stream. Background noise refers to the inherent and persistent noise in the conference environment, such as background noise from the room itself or weak noise from equipment that is not turned off. Advanced filtering algorithms, such as adaptive filtering algorithms, are employed, which can automatically adjust filtering parameters according to the characteristics of the noise to effectively remove background noise. By analyzing the noise characteristics in the synchronous multi-channel audio stream, the pattern of background noise is identified, and this pattern is subtracted from the audio stream, thereby eliminating background noise. Simultaneously, echo remnant suppression is performed. Echoes are generated when sound waves are reflected off surfaces such as the walls and ceiling of the conference room and then re-received by the microphone, interfering with the original speech signal. Using echo cancellation technology, a reference signal (such as the signal emitted by the speaker) is acquired and compared with the signal received by the microphone to estimate the characteristics of the echo path. Then, the estimated echo signal is subtracted from the microphone signal to achieve echo remnant suppression. After background noise cancellation and echo remnant suppression, purified multi-source noise data is obtained. These data, having removed inherent noise and echo interference from the environment, are purer and more accurate, providing a high-quality data foundation for subsequent steps such as voiceprint feature extraction and spectral decomposition, and noise interference level index generation. This ensures that the entire AI-based noise recognition-based digital conference power adaptive adjustment method can operate accurately and effectively.

[0073] Throughout the implementation process, attention must be paid to the microphone's installation location and orientation to ensure accurate acquisition of the corresponding type of noise data. Simultaneously, the parameter settings for timestamp alignment and frequency band normalization processing need to be adjusted according to the actual conditions and noise characteristics of the conference room to obtain the optimal synchronized multi-channel audio stream. The algorithm parameters for background noise cancellation and echo remnant suppression also need to be optimized to ensure high quality of the purified multi-source noise data. Furthermore, other noise sources that may exist in the conference room, such as projector fan noise and traffic noise outside the window, need to be considered, and their impact should be eliminated or reduced as much as possible during acquisition and processing. Through the above detailed implementation methods, comprehensive and accurate conference environment noise data can be acquired and effectively processed, providing reliable data support for subsequent noise analysis and power adjustment.

[0074] Example 2:

[0075] When constructing a convolutional-recurrent hybrid recognition model, it is necessary to clearly define the model's boundary conditions, namely, defining speech intelligibility constraints and energy conservation equations. Speech intelligibility constraints aim to ensure that the model does not substantially affect the intelligibility of conference speech during noise processing, guaranteeing that key frequency components and temporal features of the speech signal are not excessively suppressed or distorted. The energy conservation equation, as a physical constraint, requires that the total energy of the input signal and the sum of the energy of the processed output features remain essentially consistent during noise recognition and feature extraction, avoiding the creation or unnecessary loss of energy, thereby ensuring the rationality and physical accuracy of the model's processing.

[0076] A joint processor comprising a temporal convolutional module and a gated recurrent unit is constructed. The temporal convolutional module consists of multiple convolutional neural networks, each equipped with a convolutional kernel of a specific size, capable of performing sliding window-style feature extraction on the multi-dimensional noise feature spectrum of the input. This module can effectively capture the local correlations of noise signals in the temporal dimension, such as short-term features like the start time and duration of sudden noise. The gated recurrent unit consists of multiple gated neurons, including an input gate, a forget gate, and an output gate, capable of handling long-term dependencies in the noise signal. Through the gating mechanism, this unit can selectively memorize historical noise features and forget irrelevant information, thereby adapting to the continuous changes or periodic fluctuations that may occur in the noise environment of a conference.

[0077] When processing the fundamental frequency of background noise, harmonic decomposition is employed. Through signal processing techniques such as Fourier transform, the fundamental frequency signal of the background noise is decomposed into multiple harmonic components, each corresponding to a different frequency and amplitude. The distribution and intensity differences of these harmonic components reflect the physical characteristics of the noise source. For example, the harmonic structures of air conditioning noise and ventilation system noise are significantly different. Based on this, noise source characteristics are generated, providing a basis for subsequent noise source identification.

[0078] For sudden interference pulses, envelope detection processing is performed. Through rectification and low-pass filtering, the amplitude envelope curve of the pulse signal is extracted. This curve visually reflects the intensity change trend of the interference pulse, such as the gradual increase and decrease in volume of a mobile phone ringtone. Simultaneously, the duration parameter of the envelope curve is calculated to generate an impact duration feature, used to describe the duration and energy concentration of the interference pulse.

[0079] For speech aliasing data, spectral entropy is calculated. Spectral entropy is an indicator of the uncertainty of a signal's spectrum. By framing the spectrum of the aliased speech signal and estimating its power spectrum, the entropy value of each frame is calculated. When the degree of noise and speech aliasing is high, the randomness of the spectrum increases, and the spectral entropy value increases accordingly, thus generating spectral pollution characteristics, which can quantitatively reflect the degree of noise pollution in the speech signal.

[0080] Noise source features, impact duration features, and spectral pollution features are input into a coprocessor for feature fusion. The temporal convolution module first extracts spatial features from these features, capturing local patterns and details of each feature; the gated recurrent unit then processes the feature sequence from a temporal perspective, integrating historical feature information using a memory mechanism. Through nonlinear transformations of a multi-layer neural network, the complementarity and fusion of different features are achieved, ultimately outputting a noise interference level index. This index, in the form of a quantitative value, characterizes the degree of noise interference to voice communication in the current conference environment; a higher value indicates more severe interference.

[0081] During model construction, attention must be paid to the specific parameter settings of the speech intelligibility constraints. The minimum acceptable threshold for speech intelligibility should be determined based on industry standards or actual needs of the meeting scenario. The implementation of the energy conservation equation needs to be combined with specific signal processing algorithms to ensure the accuracy of energy calculations during feature extraction. The number of layers, kernel size, and number of neurons in the temporal convolutional module of the coprocessor need to be adjusted according to the complexity and feature dimensions of the noisy data. Network structure parameters can be optimized through methods such as cross-validation.

[0082] Furthermore, the accuracy of harmonic decomposition is affected by the Fourier transform window length. Appropriate window parameters must be selected based on the frequency characteristics of the background noise to avoid feature extraction deviations caused by spectral leakage. The filtering parameters in envelope detection must balance the integrity of impulse features with noise smoothing effects to prevent over-filtering from losing crucial information or under-filtering from introducing additional interference. The frame length and overlap rate in spectral entropy calculation also affect the accuracy of the results and must be reasonably selected based on the sampling rate and noise characteristics of the speech signal.

[0083] Example 3:

[0084] When performing real-time analysis of noise interference level indicators based on a dynamic threshold model, an environmental adaptability comparison table is constructed based on historical meeting noise samples. The collection of historical meeting noise samples needs to cover different meeting scenarios, such as meeting rooms of different sizes, different time periods (day or night), and different external environments (quiet or noisy). Noise data in these scenarios is collected using a distributed microphone array and preprocessed, including timestamp alignment, frequency band normalization, background noise cancellation, and echo remnant suppression, to obtain clean noise samples. Then, feature extraction is performed on these samples to obtain noise interference level indicators, while corresponding meeting environment parameters are recorded, such as the number of people in the meeting room, equipment status, and door / window status. Correlation analysis is performed between the noise interference level indicators and environmental parameters to determine the adaptability of different noise indicators in various environments, constructing an environmental adaptability comparison table. This table reflects the impact of a specific noise interference level in the corresponding environment and the system's adaptability.

[0085] After constructing the environmental adaptability comparison table, rolling time windows are used to sample the noise interference level indicators. A suitable time window length is set, for example, 10 seconds, and the noise interference level indicators are continuously sampled within each time window. The sampling frequency needs to be determined based on the rate of noise change. For rapidly changing noise, such as transient burst noise, the sampling frequency can be increased; for slowly changing steady-state noise, the sampling frequency can be appropriately decreased. After each time window, the sampled noise index values ​​are arranged in chronological order to generate an index fluctuation sequence. This sequence reflects the fluctuation of the noise interference level within the time window and can capture the short-term trend of noise change.

[0086] A double exponential smoothing algorithm is used to fit the trend of the indicator fluctuation sequence. This algorithm considers not only the current level of the sequence but also its trend changes. First, the indicator fluctuation sequence is subjected to exponential smoothing once to obtain an estimate of the current level. Then, the result of the first exponential smoothing is subjected to a second exponential smoothing to obtain a trend estimate. By continuously updating the level and trend estimates, the trend of the indicator fluctuation sequence is fitted, generating dynamic threshold parameters. These dynamic threshold parameters are not fixed but are adjusted in real time according to changes in noise, enabling a more accurate reflection of the changing trend of the current noise environment.

[0087] After obtaining the dynamic threshold parameters, their cumulative distribution is calculated based on an environmental fitness reference table. This table records the probability and impact of different noise interference levels in various environments. The dynamic threshold parameters are matched with the data in the table to calculate the cumulative probability that the noise interference level exceeds the dynamic threshold in the current environment. The cumulative distribution calculation yields a signal-to-noise ratio (SNR) degradation trend graph. This graph, with time on the horizontal axis and the degree of SNR degradation or cumulative probability on the vertical axis, visually displays the trend of SNR changes over time. For example, when the noise interference level continues to rise, the SNR degradation trend graph will show an upward trend, indicating that the noise in the meeting environment is gradually worsening, requiring timely power adjustment.

[0088] Throughout the implementation process, the collection and preprocessing of historical meeting noise samples are crucial. The diversity and representativeness of the samples directly affect the accuracy of the environmental fitness checklist; therefore, it is necessary to collect as many noise samples as possible from different scenarios. The choice of time window length needs to be determined based on the noise variation characteristics in the actual meeting scenario. If the time window is too long, rapid changes in noise may be overlooked; if the time window is too short, too much random fluctuation may be introduced, affecting the accuracy of trend fitting. The smoothing coefficient in the double exponential smoothing algorithm also needs to be set reasonably. The larger the smoothing coefficient, the faster the response to new data, but the more susceptible it is to random noise; the smaller the smoothing coefficient, the more stable the trend fitting, but the slower the response to new data.

[0089] Furthermore, when constructing the environmental adaptability comparison table, the specific needs of different meeting scenarios must be considered. For example, in important business meetings, the tolerance for noise is lower, and the threshold settings in the environmental adaptability comparison table may be more stringent; while in ordinary internal meetings, the tolerance for noise is relatively higher, and the threshold settings can be appropriately relaxed. After generating the signal-to-noise ratio degradation trend graph, the trend graph needs to be monitored and analyzed in real time. When the signal-to-noise ratio degradation trend is found to exceed the preset warning threshold, the power adjustment mechanism should be triggered in a timely manner to ensure the clarity and quality of the meeting audio.

[0090] Example 4:

[0091] When optimizing the power adjustment decision matrix using multiple objectives, a baseline gain value needs to be calculated based on the speaker response curve and the room's acoustic characteristics. The speaker response curve reflects the speaker's output gain characteristics at different frequencies. For example, a certain model of speaker may have a relatively flat gain in the mid-to-high frequency range (2kHz-8kHz), while it may experience an attenuation of about 3dB in the low frequency range (50Hz-200Hz). The room's acoustic characteristics include the room's size, shape, and interior materials. For instance, in a conference room that is 10 meters long, 8 meters wide, and 3 meters high, the reflection and attenuation of sound waves will differ significantly when the walls are made of sound-absorbing materials compared to when they are made of hard materials. By measuring the frequency response of the speakers at different locations within the room and combining this with room acoustic simulation software (such as ODEON or CATT-Acoustic) to simulate the sound field, a baseline gain value is determined. This value must ensure that, in the absence of significant noise interference, the conference audio signal can cover the entire conference room at a suitable volume, and the sound pressure level difference between different areas does not exceed 5dB.

[0092] The slope of the signal-to-noise ratio (SNR) degradation trend graph is detected to generate the adjustment demand increment. The SNR degradation trend graph plots the change in SNR under noise interference in real time, with time on the horizontal axis and SNR value on the vertical axis. For example, when an air conditioner in a conference room is suddenly turned on, the SNR may drop from 30dB to 25dB within 10 seconds; the slope of the trend graph for this time period is -0.5dB / s. The slope at each time point on the trend graph is calculated using a differential algorithm. The larger the absolute value of the slope, the faster the SNR degradation, and the greater the required power adjustment. The calculation of the adjustment demand increment requires consideration of both the sign of the slope and its absolute value. When the slope is negative and the absolute value exceeds 0.3dB / s, it is determined that an immediate power increase adjustment is needed. The increment value is proportional to the absolute value of the slope. For example, when the slope is -0.5dB / s, the adjustment demand increment is set to a 5% base gain value.

[0093] The baseline gain and adjustment demand increment are Pareto optimized using a Non-Dominated Sorting Genetic Algorithm (NSGA-II). The NSGA-II algorithm transforms multiple objectives of power adjustment (such as maximizing speech intelligibility, minimizing power consumption, and maximizing device compatibility) into a fitness function, with each objective corresponding to an optimization dimension. For example, objective 1 is to maximize speech intelligibility (evaluated by the STI-PA metric), objective 2 is to minimize system power consumption (in watts), and objective 3 is to ensure compatibility of the adjustment parameters with existing audio processing units (represented by a compatibility coefficient of 0-1). The baseline gain and adjustment demand increment are used as the genetic parameters of the initial population, with each individual representing a set of power adjustment schemes. Through genetic operations such as selection, crossover, and mutation, multiple generations of the population are generated, ultimately yielding a Pareto-optimal solution set. Each scheme in this solution set achieves a balance among multiple objectives, and no single scheme is superior to others in all objectives. For example, Scheme A may have the best speech clarity (STI-PA = 0.85), but higher power consumption (150W); Scheme B has slightly lower speech clarity (STI-PA = 0.82), but lower power consumption (120W) and higher compatibility (0.9).

[0094] The power adjustment decision matrix is ​​updated based on the Pareto optimal solution set. This matrix is ​​typically a multi-dimensional table, with rows and columns representing different noise interference levels and meeting scenario parameters. Cells store the corresponding optimized power adjustment parameters (such as gain values ​​and adjustment time intervals). For example, when the noise interference level is "medium" (corresponding to an SNR of 20-25 dB) and the number of people in the meeting room exceeds 20, a cell in the matrix specifies a gain increase of 3 dB and an adjustment time interval of 15 seconds. Then, singular value decomposition (SVD) is performed on the updated matrix, decomposing it into singular values ​​and left and right singular vectors. The magnitude of the singular values ​​reflects the importance of information in the matrix; larger singular values ​​correspond to the main features of the matrix. By analyzing the distribution of singular values, the parameter dimensions with the greatest impact on model updates are determined. For example, when the cumulative contribution of the first three singular values ​​after SVD exceeds 80%, model update instructions are generated only for the parameters corresponding to these three dimensions, such as "adjusting the kernel size of the second layer of the temporal convolution module in the convolutional recurrent hybrid recognition model from 3×3 to 5×5 to enhance the extraction capability of low-frequency noise features."

[0095] During implementation, the speaker response curve must be measured using professional acoustic measurement equipment (such as a B&K acoustic analyzer). Samples should be taken at multiple points in the conference room (such as the four corners and the center), and the average value should be taken to reduce the impact of spatial differences. Simulation of room acoustic characteristics requires accurate input of parameters such as room dimensions and material absorption coefficients. Simulation accuracy can be ensured by measuring the absorption coefficient on-site (e.g., using the impedance tube method to measure material samples). The time window setting for slope detection needs to balance real-time performance and stability, typically using a 5-10 second sliding window to avoid misjudgments caused by short-term noise fluctuations.

[0096] The population size and number of iterations in the NSGA-II algorithm need to be adjusted according to the complexity of the optimization objective. Generally, the population size is set to 50-100 and the number of iterations is set to 100-200 to ensure that the algorithm converges to a suitable Pareto front. When generating model update instructions, it is necessary to consider the changing trends of noise features. For example, when singular value decomposition indicates insufficient extraction capability of high-frequency noise features, the parameters for processing high-frequency components in the model should be adjusted accordingly, such as increasing the memory weights of high-frequency features in the gated recurrent units.

[0097] In addition, the actual hardware limitations of the conference system must be considered, such as the maximum gain range of the audio processing unit and the bandwidth of the power amplifier. When the optimized power adjustment parameters exceed the hardware capabilities, the parameters need to be trimmed or hardware upgrades should be recommended. For example, if the optimization result requires a gain increase of 10dB, but the maximum gain of the existing power amplifier is only 8dB, then the gain parameter should be adjusted to 8dB, and the need to replace it with a higher-gain power amplifier module should be noted in the model update instruction.

[0098] Through the above steps, the multi-objective parameter optimization process can ensure speech clarity while taking into account system power consumption and device compatibility, generating a reasonable power adjustment decision matrix and model update instructions. This enables the digital conference system to dynamically adjust power output according to the real-time noise environment, maintaining good conference sound quality and system efficiency in different scenarios.

[0099] Example 5:

[0100] When classifying the power levels of conference terminals based on the updated model output, the noise interference level index first needs to be analyzed by region clustering. Taking a rectangular conference room with a length of 15 meters and a width of 10 meters as an example, a distributed microphone array deploys one overhead microphone at each of the four corners and the center of the ceiling, and a desktop directional microphone is placed every 2 meters on both sides of the conference table. Mobile terminals such as mobile phones used by attendees serve as auxiliary data acquisition devices. During the meeting, if the air conditioner suddenly starts, the noise interference level index in the right area of ​​the conference room reaches 3.2, while the index in the left area, being farther from the air conditioner, is 1.8. At this point, the K-means clustering algorithm is used to divide areas with similar noise levels into different clusters. Setting the cluster size to 3, the algorithm divides the right side of the conference room into a high-interference zone (index ≥ 3.0), the middle area into a medium-interference zone (index 1.5-2.9), and the left side into a low-interference zone (index ≤ 1.4), generating a set of high-interference zones.

[0101] Subsequently, frequency band correlation calibration was performed on the dynamic gain threshold. It was assumed that in a high-interference zone, air conditioning noise is mainly concentrated in the low-frequency band (50Hz-200Hz), while keyboard typing noise is distributed in the mid-to-high frequency band (1kHz-4kHz), and the noise characteristics of these two frequency bands are somewhat correlated. By calculating the Pearson correlation coefficient of noise characteristics in different frequency bands, it was found that the correlation coefficient between low-frequency energy and mid-to-high frequency energy is 0.35, indicating a weak correlation. Based on this, environmental adaptability constraints were generated. For example, when the low-frequency noise index exceeds the threshold, the gain adjustment amplitude in the mid-to-high frequency band should be limited to within 40% of the low-frequency band adjustment amplitude to avoid sound distortion caused by mutual interference between frequency bands.

[0102] Topology mapping is performed on the high-interference partition set based on environmental adaptability constraints. In the conference room example above, the high-interference partition is located in the right-hand area, corresponding to the rectangular area with coordinates (10-15 meters, 0-10 meters) in the topology map. Considering the distribution of conference terminal locations, there are 3 desktop microphones and 2 speakers in this area. During topology mapping, the power adjustment strategy needs to be associated with the terminal locations. For example, the speaker gain in this area can be increased by 3dB, and the microphone sensitivity can be increased by 2dB, while also satisfying the frequency band correlation constraint, i.e., when the low-frequency band gain is increased by 3dB, the mid-to-high frequency band gain can be increased by a maximum of 1.2dB.

[0103] When generating the dynamic gain threshold, Gaussian mixture modeling is used for the noise interference level index. It is assumed that the noise index of a high-interference zone during a certain period follows a mixture of two Gaussian distributions: one with a mean of 3.5 (corresponding to steady-state air conditioning noise) and a standard deviation of 0.5; the other with a mean of 4.2 (corresponding to sudden equipment startup noise) and a standard deviation of 0.3. After estimating the parameters using the EM algorithm, a noise tolerance range of [2.8, 4.5] is generated. That is, when the noise index exceeds 4.5, strong power adjustment is triggered; when it is below 2.8, the default setting is restored. Differential calculation is performed on the multi-dimensional noise feature spectrum. For example, if the low-frequency energy suddenly increases by 15% and the mid-to-high frequency energy increases by 5% at a certain moment, an abnormal sound signature feature vector [15%, 5%,...] is generated. Using an adaptive threshold segmentation algorithm, regions exceeding twice the standard deviation of the local mean are marked as abrupt change regions, such as low-frequency energy abrupt change points. An initial gain threshold of 4dB is generated, and then the initial threshold is corrected for sound field uniformity. Considering that the right wall of the conference room is made of sound-absorbing material and the sound decays quickly, the initial threshold was adjusted to increase by 4.8dB to ensure that the sound pressure level is uniform at all points in the area.

[0104] When generating real-time power adjustment signals, a priority weight table is constructed based on the terminal location distribution and acoustic coverage requirements. In the aforementioned conference room, the microphone on the podium (located in the low-interference zone on the left) has a weight of 0.9, the desktop microphones of the participants (distributed in the medium-high interference zone) have a weight of 0.7, and the rear speakers have a weight of 0.6. Energy diffusion analysis is performed on the zone adjustment strategy set. Assuming that the high-interference zone needs to increase power by 5dB, the proportion of energy diffusion to the adjacent medium-interference zone is calculated to be 20% through acoustic simulation, generating adjustment intensity coefficients of [5dB, 1dB] (corresponding to high and medium interference zones, respectively). The adjustment intensity coefficients are discretized and encoded using a fuzzy state transition machine, mapping 5dB to the "emphasis" state (code 11) and 1dB to the "weak adjustment" state (code 01), generating multi-level adjustment signals. Time delay compensation processing is performed on the multi-level adjustment signals. Considering that the time for sound to travel from the audio processing unit to the right high-interference zone speaker is approximately 8ms, an 8ms time delay compensation is added to the signal to obtain the real-time power adjustment signal.

[0105] When generating the final power adjustment scheme, device compatibility matching is performed on the real-time power adjustment signal. For example, if the adjustment signal requires a 5dB gain increase for a certain type of speaker, but the maximum gain limit for that speaker is 4dB, the gain is adjusted to 4dB when generating the basic adjustment strategy, and device compatibility conflicts are recorded. Power consumption limits are checked on the basic adjustment strategy. If simultaneously increasing the power of all terminals would cause the total system power consumption to exceed the power supply's rated power (e.g., rated power is 500W, and the expected power consumption after adjustment is 550W), the adjustment amplitude is reallocated. For example, the gain of the high-interference zone is increased by 4dB, and the medium-interference zone by 0.5dB, ensuring that the total power consumption is controlled within 480W, generating a feasible adjustment instruction set. The feasible adjustment instruction set is verified through sound field reconstruction using digital twin simulation. Adjustment instructions are input into a virtual conference room model to simulate the sound field distribution, checking whether the SNR of the high-interference zone is increased to above 25dB and whether the sound pressure level difference between different areas is less than 3dB, generating the final power adjustment scheme.

[0106] When performing logical verification on the real-time power adjustment signal, a cross-comparison is performed using redundant verification channels (such as the main channel and the backup channel processing the signal simultaneously). For example, if the main channel output adjustment gain is 4dB and the backup channel output is 4.1dB, and the difference between the two is within the allowable range of 0.5dB, a verification result vector [1,1] (1 indicates consistency) is generated. The consistency of the verification result vector is evaluated. If all channel results are consistent, a verification pass flag is generated. If the verification pass flag is true, the final power adjustment scheme is executed; if it is false, such as the main channel outputting 5dB while the backup channel outputting 3dB, a safety backoff mechanism is triggered, and the previously verified effective adjustment scheme is retrieved as the backup adjustment scheme, such as increasing the gain by 3dB, and a fault alarm signal is sent to the system administrator.

[0107] During implementation, the number of clusters in the regional clustering analysis needs to be adjusted according to the size of the conference room and the noise distribution characteristics. Large conference rooms can be set to 5-7 clusters, while small conference rooms can be set to 2-3 clusters. Frequency band correlation calibration needs to be updated regularly, for example, recalculating the correlation coefficient every 30 minutes to adapt to changes in equipment usage during the meeting (such as the introduction of new noise frequency bands when the projector is turned on). The accuracy of the digital twin simulation model depends on the accuracy of the conference room's geometric parameters and material properties. Laser scanning is required to obtain the three-dimensional dimensions of the conference room, and impedance tubes are used to measure the sound absorption coefficient of the materials to ensure the reliability of the simulation results. In addition, the power consumption verification of the power adjustment scheme needs to consider the instantaneous peak power consumption of the equipment. For example, the power consumption of the speaker when it starts up may be 1.5 times that of the steady state, to avoid damage to the equipment due to instantaneous overload. Through the above specific implementation methods, precise regional adjustment of the power of the conference terminal can be achieved, maintaining the voice clarity and system stability of each area in complex noise environments.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital conference power adaptive adjustment method based on AI noise recognition, characterized in that, The method includes: The conference environment noise data is collected in real time by a distributed microphone array, and the noise data is subjected to voiceprint feature extraction and spectrum decomposition to generate a multi-dimensional noise feature spectrum. Convolutional cyclic hybrid recognition model is constructed by inputting background noise fundamental frequency, sudden interference pulses and speech aliasing data from the multi-dimensional noise feature spectrum, and generating a noise interference level index. The noise interference level index is analyzed in real time based on the dynamic threshold model, a signal-to-noise ratio degradation trend map is generated, and a power adjustment decision matrix is ​​constructed based on the signal-to-noise ratio degradation trend map. The power adjustment decision matrix is ​​optimized for multiple objectives to generate a model update instruction, and the convolutional recurrent hybrid recognition model is subjected to online incremental learning based on the model update instruction. Based on the updated model output, the power levels of the conference terminals are divided, dynamic gain thresholds are generated, and environmental adaptability constraints are applied to the dynamic gain thresholds to obtain a set of partition adjustment strategies. The audio processing unit is parameter-adapted according to the partition adjustment strategy set to generate a real-time power adjustment signal, and the real-time power adjustment signal is logically verified to generate the final power adjustment scheme.

2. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The method of collecting real-time conference environment noise data through a distributed microphone array includes: Steady-state environmental noise data was collected using a top-mounted microphone in the venue, local interference noise data was collected using a desktop directional microphone, and transient burst noise data was obtained using a mobile terminal microphone. The steady-state environmental noise data, local interference noise data, and transient burst noise data are processed by timestamp alignment and frequency band normalization to generate a synchronous multi-channel audio stream. The synchronous multi-channel audio stream is subjected to background noise cancellation and echo retention suppression to obtain purified multi-source noise data.

3. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The construction of the convolutional recurrent hybrid recognition model includes: Speech proficiency constraints and energy conservation equations are defined as boundary conditions for the model, and a joint processor containing temporal convolutional modules and gated recurrent units is constructed. Harmonic decomposition is performed on the fundamental frequency of the background noise to generate noise source features; envelope detection is performed on the sudden interference pulse to generate impact duration features; and spectral entropy is calculated on the speech aliasing data to generate spectral pollution features. The noise source characteristics, impact duration characteristics, and spectral pollution characteristics are input into the joint processor for feature fusion, and a noise interference level index is output.

4. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The real-time analysis of the noise interference level index based on the dynamic threshold model includes: An environmental fitness comparison table was constructed based on historical conference noise samples, and the noise interference level index was sampled using a rolling time window to generate an index fluctuation sequence. The index fluctuation sequence is fitted with a trend using a double exponential smoothing algorithm to generate dynamic threshold parameters. The cumulative distribution of the dynamic threshold parameters is then calculated based on the environmental fitness reference table to obtain a signal-to-noise ratio degradation trend map.

5. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The multi-objective parameter optimization of the power regulation decision matrix includes: The baseline gain value is calculated based on the speaker response curve and room acoustic characteristics, and the slope of the signal-to-noise ratio degradation trend graph is detected to generate the adjustment demand increment. The baseline gain value and the adjustment demand increment are Pareto optimized using a non-dominated sorting genetic algorithm to generate an updated power adjustment decision matrix, and the model update instruction is determined based on the matrix singular value decomposition results.

6. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The classification of conference terminal power levels based on the updated model output includes: Regional clustering analysis is performed on the noise interference level index to generate a set of high interference partitions, and frequency band correlation calibration is performed on the dynamic gain threshold to generate environmental adaptability constraints. Based on the environmental adaptability constraints, a topological mapping is performed on the set of high-interference partitions to generate a set of partition adjustment strategies.

7. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The generation of the dynamic gain threshold includes: Gaussian mixture modeling is performed on the noise interference level index to generate a noise tolerance range, and differential calculation is performed on the multi-dimensional noise feature spectrum to generate abnormal voiceprint features. The abnormal voiceprint features are marked with abrupt change regions using an adaptive threshold segmentation algorithm to generate an initial gain threshold. The initial gain threshold is then corrected for sound field uniformity to obtain a dynamic gain threshold.

8. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The generation of the real-time power adjustment signal includes: A priority weight table is constructed based on the terminal location distribution and acoustic coverage requirements, and energy diffusion analysis is performed on the partitioned adjustment strategy set to generate adjustment intensity coefficients. The adjustment intensity coefficient is discretized and encoded using a fuzzy state transition machine to generate multi-level adjustment signals. Time delay compensation processing is then applied to the multi-level adjustment signals to obtain real-time power adjustment signals.

9. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 1, characterized in that, The generation of the final power regulation scheme includes: The real-time power adjustment signal is matched for device compatibility to generate a basic adjustment strategy, and the basic adjustment strategy is verified for power consumption limitations to generate a feasible adjustment instruction set. The feasible adjustment instruction set is verified by sound field reconstruction through digital twin simulation to generate the final power adjustment scheme.

10. The digital conference power adaptive adjustment method based on AI noise recognition according to claim 9, characterized in that, The logical verification of the real-time power adjustment signal includes: The real-time power adjustment signal is cross-compared through a redundant verification channel to generate a verification result vector, and the consistency of the verification result vector is evaluated to generate a verification pass flag. If the verification pass flag is true, the final power adjustment scheme is executed; if it is false, a safety rollback mechanism is triggered and a backup adjustment scheme is generated.

Images