An artificial intelligence-based recording noise reduction method and system

Abstract

The present application relates to the technical field of recording noise reduction, and particularly relates to a recording noise reduction method and system based on artificial intelligence, which comprises: performing short-time Fourier transform on a to-be-processed recording signal to obtain observation time-frequency coefficients arranged according to frequency units and time units, and constructing a current input time-frequency window and a noise suppression sample vector; inputting the current input time-frequency window into an encoding network to extract a first time-frequency context embedding representation, outputting noise suppression filtering parameters through a noise suppression decoding network, and linearly filtering the noise suppression sample vector to obtain a noise suppression result; then constructing a decoding input time-frequency window and a reverberation suppression sample vector with a preset delay based on a time-frequency signal used for reverberation suppression calculation, outputting reverberation suppression filtering parameters through a reverberation suppression decoding network, linearly predicting the reverberation suppression sample vector to obtain a late reflection estimate, and deducting the late reflection estimate from the noise suppression result to obtain enhanced time-frequency coefficients; and performing inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal. The present application can improve the intelligibility, intelligibility and perceptual quality of the recording signal.

Description

Technical Field

[0001] This invention relates to the field of audio recording noise reduction technology, and in particular to an artificial intelligence-based audio recording noise reduction method and system. Background Technology

[0002] In applications such as far-field voice pickup, mobile terminal recording, conference capture, and online voice communication, recorded signals are often simultaneously affected by ambient background noise and indoor acoustic reflections. Background noise reduces the clarity of the target speech, while reverberation, especially late reflections, causes speech trailing, blurring, and decreased intelligibility, thus affecting recording quality and the effectiveness of subsequent applications such as speech recognition and communication enhancement. While existing deep learning-based recording enhancement methods can improve noise reduction to some extent, some high-performance methods have large model sizes and high computational resource consumption, making it difficult to meet the needs of real-time processing and edge device deployment. On the other hand, some lightweight methods, while highly efficient in noise suppression, still have limited ability to suppress late reverberation, especially in reverberation scenarios with strong long-term dependence, and are prone to the problem of insufficient noise reduction and reverberation reduction. Therefore, how to improve the noise suppression and reverberation suppression capabilities of recorded signals while maintaining low computational complexity and real-time performance has become an urgent technical problem to be solved in this field. Summary of the Invention

[0003] In view of the above technical problems, the present invention provides an artificial intelligence-based recording noise reduction method and system to solve the problem that existing lightweight recording enhancement schemes are difficult to simultaneously achieve background noise suppression, late reverberation suppression and real-time processing performance in complex acoustic environments, thereby improving the clarity, intelligibility and perceived quality of the recording signal.

[0004] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0005] According to one aspect of the present invention, an artificial intelligence-based method for recording noise reduction is proposed, the method comprising the steps of: A short-time Fourier transform is performed on the recording signal to be processed to obtain the observed time-frequency coefficients arranged by frequency and time units, and the current input time-frequency window and noise suppression sample vector are constructed based on the observed time-frequency coefficients. The current input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the first time-frequency context embedding representation, and the noise suppression decoding network in the artificial intelligence enhancement model outputs noise suppression filtering parameters according to the first time-frequency context embedding representation; The noise suppression sample vector is linearly filtered using the noise suppression filtering parameters to obtain the noise suppression result. A decoding input time-frequency window and a reverberation suppression sample vector are constructed based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. The decoding input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the second time-frequency context embedding representation. The reverberation suppression decoding network in the artificial intelligence enhancement model outputs reverberation suppression filtering parameters based on the second time-frequency context embedding representation. The reverberation suppression filtering parameters are then used to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate. The late reflection estimate is then subtracted from the noise suppression result to obtain the enhancement time-frequency coefficients. Perform an inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

[0006] Furthermore, the recording signal to be processed is obtained by superimposing the reverberant speech formed by convolving the target speech with the acoustic impulse response and the background noise. The acoustic impulse response is divided into an early component containing direct sound and early reflections and a late component containing late reflections. The noise-reduced recording signal is used to retain the speech information corresponding to the early component and suppress the background noise and the reverberation information corresponding to the late component.

[0007] Furthermore, the current input time-frequency window is composed of the observation time-frequency coefficients of multiple consecutive frames arranged around the current time unit and all frequency units; the noise suppression sample vector is composed of the observation time-frequency coefficients of the current frequency unit corresponding to the current time unit and the observation time-frequency coefficients of its adjacent time series; the reverberation suppression sample vector is composed of the continuous historical time-frequency coefficients of the current frequency unit corresponding to the time-frequency signal used for reverberation suppression calculation after a delay relative to the current time unit; the vector length of the reverberation suppression sample vector is greater than the vector length of the noise suppression sample vector, so that the reverberation suppression decoding network represents a time correlation longer than that of noise suppression.

[0008] Furthermore, the artificial intelligence enhancement model also includes a masking decoding network, and the encoding network is a shared encoding network; the method further includes dividing the observed time-frequency coefficients into low-frequency intervals and high-frequency intervals according to preset boundary frequency units, extracting complex time-frequency features from the low-frequency intervals, extracting equivalent rectangular bandwidth features obtained by amplitude spectrum downsampling from the high-frequency intervals, and inputting the complex time-frequency features and the equivalent rectangular bandwidth features into the shared encoding network so that the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network can output corresponding enhancement parameters respectively.

[0009] Furthermore, the masking decoding network is used to output real-valued masking parameters for the high-frequency range. These real-valued masking parameters are predicted in the equivalent rectangular bandwidth domain and then interpolated back to the short-time Fourier transform domain, where they are applied to the corresponding high-frequency observation time-frequency coefficients. The noise suppression decoding network is used to output complex linear filtering parameters for the low-frequency range. The reverberation suppression decoding network is used to output reverberation suppression filtering parameters for the low-frequency range, where the reverberation suppression filtering parameters are complex linear prediction parameters. The enhancement result for the low-frequency range is obtained by subtracting the noise suppression result formed by the complex linear filtering parameters applied to the noise suppression sample vector from the late reflection estimate formed by the complex linear prediction parameters applied to the reverberation suppression sample vector. The enhancement result for the high-frequency range and the enhancement result for the low-frequency range are concatenated to form the enhanced time-frequency coefficients.

[0010] Furthermore, the shared coding network generates an embedded representation covering all frequency units in a single forward computation. The noise suppression decoding network and the reverberation suppression decoding network output corresponding complex enhancement parameters according to the frequency units. The noise suppression sample vector and the reverberation suppression sample vector are both composed of continuous time-series time-frequency coefficients on the same frequency unit, so as to complete the enhancement operation in the low-frequency range without introducing cross-band filtering.

[0011] Furthermore, the time-frequency signal used for reverberation suppression calculation is an intermediate time-frequency signal formed from the first-stage enhancement result. The method employs a two-step inference approach to perform enhancement processing. In the first step, the observed time-frequency coefficients are enhanced using the shared coding network, the masking decoding network, and the noise suppression decoding network to obtain the intermediate time-frequency signal. In the second step, the decoding input time-frequency window and the reverberation suppression sample vector are constructed based on the intermediate time-frequency signal. The reverberation suppression filter parameters are output using the shared coding network and the reverberation suppression decoding network. The reverberation suppression filter parameters are then applied to the reverberation suppression sample vector to obtain an intermediate late reflection estimate. The intermediate late reflection estimate is subtracted from the intermediate time-frequency signal to obtain the final enhanced time-frequency coefficients. In the second step, the masking decoding network and the noise suppression decoding network are not enabled, and reverberation suppression is performed in the low-frequency range, while the first-stage enhancement result is maintained in the high-frequency range.

[0012] Furthermore, the AI ​​enhancement model is trained as follows: a training set is constructed based on pairs of clean speech samples and noisy reverberant speech samples; the shared encoding network, the masking decoding network, and the noise suppression decoding network are initialized using pre-trained enhancement model parameters; in the early stage of training, the parameters of the other networks except the reverberation suppression decoding network are fixed, and only the parameters of the reverberation suppression decoding network are updated; after the reverberation suppression decoding network completes initial convergence, the shared encoding network, the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network are jointly trained end-to-end; the joint training adopts an objective function that includes spectral consistency loss and multi-resolution loss, wherein the spectral consistency loss calculates the amplitude difference after perceptually compressing the spectral amplitude of clean speech and enhanced speech, and constrains the phase consistency; the multi-resolution loss constrains the compressed spectral amplitude difference and phase consistency at multiple analysis window scales respectively, so as to improve the perceptual quality and reverberation suppression capability of the noise-reduced recording signal.

[0013] Furthermore, the time-frequency signal used for reverberation suppression calculation is the time-frequency signal corresponding to the observed time-frequency coefficients. The decoding input time-frequency window is constructed by delaying the observed time-frequency coefficients in time. The noise suppression decoding network outputs the noise suppression filtering parameters according to the first time-frequency context embedding representation extracted from the current input time-frequency window. The reverberation suppression decoding network outputs the reverberation suppression filtering parameters according to the second time-frequency context embedding representation extracted from the decoding input time-frequency window. The noise suppression result and the late reflection estimate are calculated in the same forward inference, so as to simultaneously complete background noise suppression and late reverberation suppression in a single enhancement calculation.

[0014] According to another aspect of the present invention, an artificial intelligence-based audio recording noise reduction system is provided, comprising: The time-frequency transformation and input construction module is used to perform short-time Fourier transform on the recording signal to be processed, obtain the observed time-frequency coefficients arranged by frequency unit and time unit, and construct the current input time-frequency window and noise suppression sample vector based on the observed time-frequency coefficients; The noise suppression parameter estimation module is used to input the current input time-frequency window into the encoding network of the artificial intelligence enhancement model, extract the first time-frequency context embedding representation, and use the noise suppression decoding network in the artificial intelligence enhancement model to output noise suppression filtering parameters according to the first time-frequency context embedding representation; The noise suppression execution and reverberation input construction module is used to perform linear filtering on the noise suppression sample vector using the noise suppression filtering parameters to obtain the noise suppression result, and to construct a decoding input time-frequency window and a reverberation suppression sample vector based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. The reverberation suppression parameter estimation and late reflection cancellation module is used to input the decoded input time-frequency window into the encoding network in the artificial intelligence enhancement model, extract the second time-frequency context embedding representation, and use the reverberation suppression decoding network in the artificial intelligence enhancement model to output reverberation suppression filtering parameters according to the second time-frequency context embedding representation. The module then uses the reverberation suppression filtering parameters to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate, and subtracts the late reflection estimate from the noise suppression result to obtain the enhancement time-frequency coefficients. The time-domain reconstruction output module is used to perform an inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

[0015] The technical solution of the present invention has the following beneficial effects: Compared with existing technologies, this invention can more effectively balance noise suppression and reverberation suppression during the recording enhancement process, especially improving the suppression effect on late reverberation, thereby improving the speech clarity, perceived quality and overall intelligibility of the enhanced recording; at the same time, this invention still maintains good computational efficiency and real-time processing capabilities, and is suitable for resource-constrained devices or real-time speech processing scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of an artificial intelligence-based audio noise reduction method as described in the embodiments of this specification; Figure 2 This is a structural block diagram of an artificial intelligence-based recording noise reduction system as described in the embodiments of this specification. Detailed Implementation

[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention may be practiced with one or more of these specific details omitted, or other methods, components, systems, steps, etc., may be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of the invention.

[0018] Furthermore, the accompanying drawings are merely illustrative of the invention. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0019] This invention provides an artificial intelligence-based method for recording noise reduction. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating an artificial intelligence-based audio noise reduction method according to an embodiment of the present invention. The method is executed in an Android all-in-one advertising screen, particularly a battery-powered Android all-in-one advertising screen terminal. Specifically, the method may include the following steps S101-S105: In step S101, a short-time Fourier transform is performed on the recording signal to be processed to obtain the observed time-frequency coefficients arranged by frequency unit and time unit, and the current input time-frequency window and noise suppression sample vector are constructed based on the observed time-frequency coefficients.

[0020] In step S102, the current input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the first time-frequency context embedding representation, and the noise suppression decoding network in the artificial intelligence enhancement model outputs noise suppression filtering parameters according to the first time-frequency context embedding representation.

[0021] In step S103, the noise suppression sample vector is linearly filtered using the noise suppression filtering parameters to obtain the noise suppression result. A decoding input time-frequency window and a reverberation suppression sample vector are constructed based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. In step S104, the decoding input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the second time-frequency context embedding representation. The reverberation suppression decoding network in the artificial intelligence enhancement model outputs reverberation suppression filtering parameters based on the second time-frequency context embedding representation. The reverberation suppression filtering parameters are used to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate. The late reflection estimate is then subtracted from the noise suppression result to obtain the enhancement time-frequency coefficients. In step S105, an inverse short-time Fourier transform is performed on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

[0022] The recording signal to be processed is obtained by superimposing the reverberant speech formed by the convolution of the target speech with the acoustic impulse response and the background noise. The acoustic impulse response is divided into an early component containing direct sound and early reflections and a late component containing late reflections. The noise-reduced recording signal obtained thereafter is used to retain the speech information corresponding to the early component and suppress the background noise and the reverberation information corresponding to the late component.

[0023] The current input time-frequency window is composed of the observed time-frequency coefficients of multiple consecutive frames arranged around the current time unit and all frequency units; the noise suppression sample vector is composed of the observed time-frequency coefficients of the current frequency unit in the current time unit and the observed time-frequency coefficients of its adjacent time series; the reverberation suppression sample vector is composed of the continuous historical time-frequency coefficients of the current frequency unit corresponding to the time-frequency signal used for reverberation suppression calculation, after a delay relative to the current time unit; the vector length of the reverberation suppression sample vector is greater than the vector length of the noise suppression sample vector, so that the reverberation suppression decoding network represents a time correlation longer than that of noise suppression.

[0024] For explanation, in this embodiment, the recording signal to be processed can be regarded as a degraded recording formed after the target speech propagates in the actual sound field and is acquired by a single channel. It includes both reverberation components from room acoustic reflections and external background noise components. The corresponding time-domain signal model can be written as: ; Where x(n) represents the target speech, y(n) represents the recording signal to be processed, and h(n) represents the acoustic impulse response. This indicates the early component, which includes direct sound and early reflections. Let represent the late component, characterize the late reflection, v(n) represent the background noise, and * denote linear convolution. The goal of enhancement processing is to recover the effective speech components, mainly direct sound and early reflections, from degraded recordings, while suppressing background noise and late reverberation. The corresponding relationship can be expressed as: ; in, This represents the enhanced recording signal, and G represents the enhancement mapping relationship. With this setting, the enhancement result is not only used to reduce noise interference, but also to mitigate the trailing and blurring caused by late reflections, making the output recording closer to the target result of preserving early effective speech information.

[0025] In step S101, after performing a short-time Fourier transform on the recorded signal to be processed, the observed time-frequency coefficients are obtained, arranged by frequency and time units. The frequency unit can be denoted as... The time unit can be denoted as The corresponding observation time-frequency coefficients are denoted as Based on this, a current input time-frequency window is constructed. This window consists of multiple consecutive time frames near the current time unit and all frequency units, and is used to provide the coding network with cross-time and cross-frequency context information. This time-frequency window can be represented as: ; in, This indicates the number of frames covered by the time window. The current input time-frequency window is organized around the current time unit. Depending on real-time requirements, it can use only the current and historical frames, or it can include some future frames if delay is allowed. With this form, the coding network can obtain time-frequency context information covering multiple frequency units and multiple time units in a single forward computation, thus providing a basis for subsequent filter parameter estimation.

[0026] Corresponding to the current input time-frequency window, a noise suppression sample vector is also constructed. The noise suppression sample vector consists of the observed time-frequency coefficients of the current frequency unit in the current time unit and the observed time-frequency coefficients of its adjacent time series, belonging to a local time series vector for a single frequency channel. Under causal processing conditions, the noise suppression sample vector can be written as: ; Where N represents the length of the noise-suppressed sample vector, Indicates causal factors, when =0 corresponds to causal filtering. This vector is used to characterize the short-term correlation near the current frequency cell to meet the needs of noise suppression tasks for utilizing local temporal features. Since noise suppression usually focuses on modeling the current speech-noise mixture relationship, using a relatively short sample vector can achieve effective results while taking real-time performance into account.

[0027] In step S102, the current input time-frequency window is input into the encoding network of the AI ​​enhancement model, and after encoding, a first time-frequency context embedding representation is obtained. The encoding network jointly extracts multi-frame and multi-frequency features from the input window to generate embedding features suitable for subsequent decoding networks. The noise suppression decoding network outputs corresponding noise suppression filter parameters for each frequency unit based on the first time-frequency context embedding representation. These noise suppression filter parameters can be represented as a complex filter vector: ; in, Represents the coding network, This represents a noise suppression decoding network. This represents the noise suppression filter parameters output at frequency unit f and time unit t. Here, the filter parameters are output by frequency unit, but the encoding process utilizes a time-frequency window encompassing all frequency units. Therefore, it retains the frequency-dependent parameterized output format while leveraging full-frequency context information to improve parameter estimation accuracy.

[0028] In step S103, linear filtering is performed on the noise suppression sample vector using the noise suppression filtering parameters to obtain the noise suppression result. The corresponding calculation can be written as: ; in, This indicates the conjugate transpose. This represents the noise suppression result. Essentially, this calculation method performs complex linear filtering on the current frequency channel within the short-time Fourier transform domain. Compared to methods that only use real-valued gain for amplitude adjustment, it can utilize both amplitude and phase information simultaneously, improving the ability to recover target speech under complex noise conditions. After the noise suppression result is generated, the decoding input time-frequency window and reverberation suppression sample vector are constructed based on the time-frequency signal used for reverberation suppression calculation. The time-frequency signal used for reverberation suppression calculation can be the intermediate time-frequency signal corresponding to the noise suppression result obtained in the previous step, so as to reduce the interference of background noise on late reflection estimation before entering the reverberation suppression stage, thereby improving the prediction quality of the reverberation suppression filter parameters.

[0029] The reverberation suppression sample vector is composed of historical time-frequency coefficients with a preset delay relative to the current time unit, to characterize the long-term dependency information corresponding to late reflections. This sample vector can be written as: ; Where D represents the preset delay, and M represents the length of the reverberation suppression sample vector. The preset delay is used to distinguish direct sound and early reflections from subsequent late reflections in time, facilitating the model's focus on estimating the late reverberation components that have a greater impact on speech intelligibility. In practical settings, the delay parameter can be determined based on the characteristics of the reverberation environment, typically corresponding to a time boundary of approximately 50ms. Since the formation of late reflections is related to historical information over a longer time span, the length of the reverberation suppression sample vector is usually greater than the length of the noise suppression sample vector, i.e., M>N. With this setting, the reverberation suppression decoding network can characterize the temporal correlation longer than the noise suppression task, making it more suitable for modeling and predicting late reflections.

[0030] In step S104, the decoding input time-frequency window constructed based on the intermediate time-frequency signal is input to the encoding network to extract the second time-frequency context embedding representation. Then, the reverberation suppression decoding network outputs the reverberation suppression filter parameters according to the second time-frequency context embedding representation. The corresponding filter parameters can be written as follows: ; in, This indicates a reverberation suppression decoding network. This represents the decoded input time-frequency window constructed based on the intermediate time-frequency signal and containing a time delay relationship. This represents the reverberation suppression filter parameter. This filter parameter is used to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate. The corresponding calculation can be written as: ; in, This represents the late-stage reflection estimator. This represents the enhancement time-frequency coefficient. The process can be understood as follows: first, an intermediate time-frequency result is obtained through noise suppression filtering; then, late reflections are predicted using time-delay history information; and finally, the predicted late reflection components are subtracted from the intermediate time-frequency result. This allows the enhancement time-frequency coefficient to retain as much direct sound and early reflections as possible, while weakening the trailing effect caused by late reverberation. Because reverberation suppression uses a linear prediction form, its target is the historical time-frequency sequence with introduced delays. Therefore, compared to directly performing local enhancement at the current moment, it is more suitable for characterizing the long-term correlation structure in reverberation scenarios.

[0031] In step S105, an inverse short-time Fourier transform is performed on the enhanced time-frequency coefficients to recover the time-domain denoised recording signal. The inverse transform process can use the same window function and frame shift parameters as the short-time Fourier transform to ensure good stability and continuity of the time-frequency domain processing results during time-domain reconstruction. The output denoised recording signal retains the effective information of the target speech while suppressing background noise and late reverberation, thus it can be used in subsequent scenarios such as voice communication, conference recording enhancement, far-field sound pickup improvement, and speech recognition front-end processing.

[0032] In this embodiment, the current input time-frequency window, noise suppression sample vector, decoding input time-frequency window, and reverberation suppression sample vector each serve different functions. The current input time-frequency window focuses on providing full-frequency, multi-frame context information, facilitating the encoding network to learn the time-frequency structure near the current moment; the noise suppression sample vector focuses on constructing the short-time local filter input for the current frequency unit; the decoding input time-frequency window focuses on providing the reverberation suppression decoding network with a pre-enhanced time-frequency context; and the reverberation suppression sample vector focuses on characterizing late reflections by introducing a delayed historical sequence. Since the length of the reverberation suppression sample vector is greater than that of the noise suppression sample vector, the model can adapt to two different acoustic degradation mechanisms—short-time noise suppression and long-time reverberation suppression—in the same enhancement process, thereby improving the overall enhancement effect in complex recording scenarios while maintaining single-channel processing and a low computational burden.

[0033] In one embodiment, the artificial intelligence enhancement model further includes a masking decoding network, and the encoding network is a shared encoding network; the method further includes dividing the observed time-frequency coefficients into low-frequency intervals and high-frequency intervals according to preset boundary frequency units, extracting complex time-frequency features from the low-frequency intervals, extracting equivalent rectangular bandwidth features obtained by amplitude spectrum downsampling from the high-frequency intervals, and inputting the complex time-frequency features and the equivalent rectangular bandwidth features into the shared encoding network so that the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network can output corresponding enhancement parameters respectively.

[0034] The masking decoding network outputs real-valued masking parameters for the high-frequency range. These real-valued masking parameters are predicted in the equivalent rectangular bandwidth domain and then interpolated back to the short-time Fourier transform domain, where they are applied to the corresponding high-frequency observation time-frequency coefficients. The noise suppression decoding network outputs complex linear filtering parameters for the low-frequency range. The reverberation suppression decoding network outputs reverberation suppression filtering parameters for the low-frequency range, where the reverberation suppression filtering parameters are complex linear prediction parameters. The enhancement result for the low-frequency range is obtained by subtracting the noise suppression result formed by the complex linear filtering parameters applied to the noise suppression sample vector from the late reflection estimate formed by the complex linear prediction parameters applied to the reverberation suppression sample vector. The enhancement result for the high-frequency range is concatenated with the enhancement result for the low-frequency range to form the enhanced time-frequency coefficients.

[0035] The shared coding network generates an embedded representation covering all frequency units in a single forward computation. The noise suppression decoding network and the reverberation suppression decoding network output corresponding complex enhancement parameters according to the frequency units. The noise suppression sample vector and the reverberation suppression sample vector are both composed of continuous time-series time-frequency coefficients on the same frequency unit, so as to complete the enhancement operation in the low-frequency range without introducing cross-band filtering.

[0036] The time-frequency signal used for reverberation suppression calculation is an intermediate time-frequency signal formed from the first-stage enhancement result. The method employs a two-step inference approach to perform enhancement processing. In the first step, the observed time-frequency coefficients are enhanced using the shared coding network, the masking decoding network, and the noise suppression decoding network to obtain the intermediate time-frequency signal. In the second step, the decoding input time-frequency window and the reverberation suppression sample vector are constructed based on the intermediate time-frequency signal. The reverberation suppression filter parameters are output using the shared coding network and the reverberation suppression decoding network. The reverberation suppression filter parameters are then applied to the reverberation suppression sample vector to obtain an intermediate late reflection estimate. The intermediate late reflection estimate is subtracted from the intermediate time-frequency signal to obtain the final enhanced time-frequency coefficients. In the second step, the masking decoding network and the noise suppression decoding network are not enabled. The second step performs reverberation suppression for the low-frequency range, while maintaining the first-stage enhancement result for the high-frequency range.

[0037] The AI-enhanced model is trained as follows: a training set is constructed based on pairs of clean speech samples and noisy reverberant speech samples; the shared encoding network, the masking decoding network, and the noise suppression decoding network are initialized using pre-trained enhancement model parameters; in the early stages of training, the parameters of the remaining networks except the reverberation suppression decoding network are fixed, and only the parameters of the reverberation suppression decoding network are updated; after the reverberation suppression decoding network completes initial convergence, the shared encoding network, the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network are jointly trained end-to-end; the joint training adopts an objective function that includes spectral consistency loss and multi-resolution loss, wherein the spectral consistency loss calculates the amplitude difference after perceptually compressing the spectral amplitude of clean speech and enhanced speech, and constrains the phase consistency; the multi-resolution loss constrains the compressed spectral amplitude difference and phase consistency at multiple analysis window scales respectively, so as to improve the perceptual quality and reverberation suppression capability of the noise-reduced recording signal.

[0038] The time-frequency signal used for reverberation suppression calculation is the time-frequency signal corresponding to the observed time-frequency coefficients. The decoding input time-frequency window is constructed by delaying the observed time-frequency coefficients in time. The noise suppression decoding network outputs the noise suppression filtering parameters according to the first time-frequency context embedding representation extracted from the current input time-frequency window. The reverberation suppression decoding network outputs the reverberation suppression filtering parameters according to the second time-frequency context embedding representation extracted from the decoding input time-frequency window. The noise suppression result and the late reflection estimate are calculated in the same forward inference, so as to simultaneously complete background noise suppression and late reverberation suppression in a single enhancement calculation.

[0039] As an explanation, in this embodiment, the artificial intelligence enhancement model further introduces a masking decoding network and a reverberation suppression decoding network on the basis of the original noise suppression branch, and uses a shared coding network to uniformly model the input features of different frequency bands. The role of the shared coding network is to jointly represent the time-frequency context near the current time, and then provide the corresponding enhancement parameters to the masking decoding network, noise suppression decoding network, and reverberation suppression decoding network respectively. With this setting, different enhancement mechanisms are used in the high-frequency and low-frequency ranges. The high-frequency range mainly uses real-value masking, while the low-frequency range mainly uses complex linear filtering and complex linear prediction, thus balancing computational efficiency and enhancement performance. For the high-frequency range, the equivalent rectangular bandwidth feature is used as input. This feature is a downsampled form of the short-time Fourier transform amplitude spectrum, which can represent high-frequency envelope information with low complexity. For the low-frequency range, complex time-frequency features are used to retain the joint information of amplitude and phase, which is used to process periodic noise, speech harmonics, and reverberation tails more finely. After dividing the observed time-frequency coefficients into high-frequency and low-frequency ranges according to the preset boundary frequency units, the frequency units satisfy The portion is designated as the high-frequency range, satisfying... The portion is designated as the low-frequency range, in which This serves as the boundary frequency unit. The shared coding network generates an embedding representation covering all frequency units in a single forward computation. Each decoding network then outputs parameters according to the frequency unit, achieving differentiated enhancements for different frequency bands while maintaining a unified time-frequency context modeling capability. Since this embodiment is limited to not introducing cross-band filtering, the enhancement operations on each frequency unit in the low-frequency range are all completed based on continuous time-series time-frequency coefficients of the same frequency unit, corresponding to... This processing method avoids the additional complexity caused by cross-frequency coupling.

[0040] The enhancement parameters output by the masking decoding network are real-valued masking parameters in the high-frequency range, the enhancement parameters output by the noise suppression decoding network are complex linear filtering parameters in the low-frequency range, and the enhancement parameters output by the reverberation suppression decoding network are complex linear prediction parameters in the low-frequency range. For the high-frequency range, the real-valued masking parameters are directly applied to the corresponding high-frequency observation time-frequency coefficients, and the enhancement result can be written as: ; in, Represents the real-valued masking parameter. This represents the observed time-frequency coefficients. After being predicted within the equivalent rectangular bandwidth domain, this real-valued masking parameter is interpolated back to the short-time Fourier transform domain and applied to the high-frequency observed time-frequency coefficients. This approach is used because noise in the high-frequency range mainly manifests as perturbation of the speech envelope, and real-valued masking can achieve high-efficiency suppression. For the low-frequency range, enhancement does not employ simple real-valued masking but instead uses deep filtering. Deep filtering predicts the enhancement result using a linear combination of multiple short-time Fourier transform coefficients, and its expression is: ; in, Represents the parameters of a complex linear filter. This represents its i-th element. Indicates causal factors, Representing complex conjugate. When represented in vector form, the noise suppression sample vector and the complex linear filtering relationship in the low-frequency range are respectively: ; ; in, This indicates the conjugate transpose. This represents the noise suppression results in the low-frequency range. By employing complex linear filtering instead of real-valued masking, it is possible to utilize both amplitude and phase information simultaneously, thereby improving the enhancement capability in the low-frequency range when transient noise, periodic components, and speech harmonics are mixed.

[0041] The reverberation suppression decoding network outputs complex linear prediction parameters for the low-frequency range, estimating late reflections based on delayed historical samples. The reverberation suppression sample vector consists of continuous historical time-frequency coefficients delayed relative to the current time unit, and its expression is: ; Where D represents the preset delay, and M represents the length of the reverberation suppression sample vector. The preset delay is used to separate the direct sound and early reflections from the late reflections in time, making it easier for the reverberation suppression decoding network to focus on estimating the late reflection components. The length of the reverberation suppression sample vector is greater than the length of the noise suppression sample vector, i.e., M>N is usually satisfied, because late reverberation has a longer time dependence and requires a longer filter model. The traditional expression for reverberation suppression can be written as: ; In this embodiment, the reverberation suppression decoding network takes the embedded representation of the shared coding network output as input and predicts complex linear prediction parameters. Then, linear prediction is performed on the reverberation suppression sample vector to obtain the late reflection estimate. Therefore, the final enhancement result in the low-frequency range can be expressed as the noise suppression result minus the late reflection estimate, i.e.: ; When the intermediate time-frequency signal is used as the input for reverberation suppression in the low-frequency range, the reverberation suppression sample vector in the formula can also be composed of the intermediate time-frequency signal. Its function is still to estimate late reflections based on the delay history information and then subtract them from the intermediate time-frequency signal. After this processing, the low-frequency range can simultaneously achieve background noise suppression and late reverberation suppression without the need to introduce cross-frequency filtering.

[0042] The enhancement results from the high-frequency and low-frequency regions are concatenated in the time-frequency domain to form complete enhanced time-frequency coefficients, and then an inverse short-time Fourier transform is performed to recover the time-domain enhanced recording. The masking enhancement method is retained in the high-frequency region because high-frequency enhancement is relatively easier to achieve than low-frequency enhancement, while most speech energy is concentrated in the low-frequency region. Introducing complex linear filtering and complex linear prediction in the low-frequency region is more beneficial to improving the overall enhancement gain. Accordingly, continuing to use real-valued masking in the high-frequency region can control model complexity, while in the low-frequency region, complex linear filtering handles noise suppression, and complex linear prediction handles late reverberation suppression, functionally distinguishing the two types of degradation mechanisms. The shared coding network encodes the full-frequency time-frequency context through a single forward computation, and then different decoding networks generate the required parameters for high and low frequencies respectively. This structure allows the model to cover multiple enhancement targets while maintaining a lightweight design.

[0043] When employing a two-step inference approach, the first step performs a first-stage enhancement on the observed time-frequency coefficients. The masking decoding network and the noise suppression decoding network participate in the computation, while the reverberation suppression decoding network is not activated in this stage. In the first step, the intermediate time-frequency signal in the low-frequency range is obtained from the complex linear filter parameters generated by the noise suppression decoding network, and its expression is: ; ; In the high-frequency range, the observed time-frequency coefficients are enhanced based on real-valued masking parameters, and then concatenated with the intermediate time-frequency signal in the low-frequency range to form the first-stage enhancement result. In the second step, the intermediate time-frequency signal is formed using the first-stage enhancement result. Based on this intermediate time-frequency signal, a decoding input time-frequency window and a reverberation suppression sample vector are constructed. Then, the shared coding network and the reverberation suppression decoding network output the reverberation suppression filter parameters to complete the estimation and subtraction of late reflections. The expression is as follows: ; ; In the two-step inference approach, the masking decoding network and noise suppression decoding network are not enabled in the second step. Reverberation suppression is only performed on the low-frequency range, while the enhancement results from the first stage are retained in the high-frequency range. This approach decouples noise suppression and reverberation suppression along the inference path. The reverberation suppression decoding network deals with the already denoised intermediate time-frequency signal, which helps improve the accuracy of late reflection estimation. However, this comes at the cost of requiring two forward calculations, which increases inference latency and computational load accordingly.

[0044] In the single-forward synchronous inference method, noise suppression and reverberation suppression are performed in the same enhancement calculation. In this case, the noise suppression decoding network represents the output complex linear filter parameters based on the first time-frequency context embedding extracted from the current input time-frequency window, and the reverberation suppression decoding network represents the output complex linear prediction parameters based on the second time-frequency context embedding extracted from the decoded input time-frequency window constructed with a time delay. The low-frequency enhancement expression under the synchronous inference method is: ; ; ; Correspondingly, the high-frequency range is still enhanced using real-valued masking, and the results are concatenated with the low-frequency enhancement results to obtain the final enhanced time-frequency coefficients. The synchronous inference method requires only one forward computation, thus being closer to the original lightweight framework in terms of latency. However, the reverberation suppression decoding network directly predicts complex linear prediction parameters from noisy input, which is more difficult than predicting from intermediate time-frequency signals. Furthermore, due to the end-to-end joint training of the entire model, coupling may occur between the complex linear filtering parameters and the complex linear prediction parameters; the reverberation suppression branch may also have noise suppression effects, and vice versa. This approach is suitable for applications with higher requirements for the latency of a single enhancement computation.

[0045] In terms of training, the AI-enhanced model can construct a training set using pairs of clean speech samples and noisy reverberant speech samples, with a sampling rate of 48kHz. The shared coding network, masking decoding network, and noise suppression decoding network can be initialized using pre-trained enhancement model parameters, while the reverberation suppression decoding network uses a new output layer to adapt to the reverberation suppression filter parameter dimension. Initially, the parameters of all networks except the reverberation suppression decoding network are fixed, and only the reverberation suppression decoding network is trained to establish a stable late reflection estimation capability. After the reverberation suppression decoding network achieves initial convergence, the shared coding network, masking decoding network, noise suppression decoding network, and reverberation suppression decoding network are then jointly trained end-to-end. In implementation, the noise suppression filter length can be set to N=5, and the reverberation suppression filter length can be set to M=10. The structural difference between the reverberation suppression decoding network and the noise suppression decoding network can be reflected in the final output layer dimension. This setup maintains the initialization advantages of the original lightweight framework while preventing the newly added reverberation suppression branch from being significantly disturbed by other converged modules in the early stages of training.

[0046] The joint training employs an objective function composed of spectral consistency loss and multi-resolution loss. The spectral consistency loss constrains the amplitude and phase consistency between enhanced and clean speech at a single short-time Fourier transform scale, and its expression is: ; in, Representing complex numbers The phase, c=0.6, is the perceptual compression factor used to characterize loudness perception characteristics. The first term in this loss primarily constrains the difference in compressed spectral amplitude, guiding the training of the masked decoding network; the second term simultaneously constrains amplitude and phase consistency, guiding the complex linear filtering and complex linear prediction branches in the low-frequency range. The multi-resolution loss is used to simultaneously constrain the difference in compressed spectral amplitude and phase consistency between enhanced and clean speech at multiple analysis window scales, and its expression is: ; in, and Let these represent the time-frequency representations of clean speech and enhanced speech at the i-th short-time Fourier transform scale, respectively. The analysis window length can be taken as... Multi-resolution loss enhances the time-frequency structure of the results through multi-scale constraints, which helps improve perception quality and reverberation suppression.

[0047] The key structural feature of this embodiment lies in incorporating real-valued masking enhancement in the high-frequency range, complex linear filtering enhancement in the low-frequency range, and complex linear prediction déresonance in the low-frequency range into the same shared coding framework. Furthermore, it explicitly models late reflections by adding a reverberation suppression decoding network and a delayed input mechanism. This model maintains the low-complexity advantages of using equivalent rectangular bandwidth features and real-valued masking in the high-frequency range, while utilizing longer history vectors and complex parameters to handle the superposition and degradation of noise and reverberation in the low-frequency range. Thus, it achieves joint enhancement for real-time applications without introducing cross-band filtering.

[0048] Based on the same line of thought, such as Figure 2 As shown, an artificial intelligence-based audio recording noise reduction system is provided, comprising: The time-frequency transformation and input construction module 201 is used to perform short-time Fourier transform on the recording signal to be processed, obtain the observed time-frequency coefficients arranged by frequency unit and time unit, and construct the current input time-frequency window and noise suppression sample vector based on the observed time-frequency coefficients; The noise suppression parameter estimation module 202 is used to input the current input time-frequency window into the encoding network in the artificial intelligence enhancement model, extract the first time-frequency context embedding representation, and use the noise suppression decoding network in the artificial intelligence enhancement model to output noise suppression filtering parameters according to the first time-frequency context embedding representation; The noise suppression execution and reverberation input construction module 203 is used to perform linear filtering on the noise suppression sample vector using the noise suppression filtering parameters to obtain the noise suppression result, and to construct a decoding input time-frequency window and a reverberation suppression sample vector based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. The reverberation suppression parameter estimation and late reflection cancellation module 204 is used to input the decoded input time-frequency window into the encoding network in the artificial intelligence enhancement model, extract the second time-frequency context embedding representation, and use the reverberation suppression decoding network in the artificial intelligence enhancement model to output reverberation suppression filtering parameters according to the second time-frequency context embedding representation. The reverberation suppression filtering parameters are then used to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate. Finally, the late reflection estimate is subtracted from the noise suppression result to obtain the enhancement time-frequency coefficients. The time-domain reconstruction output module 205 is used to perform an inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

[0049] Compared with existing technologies, this system can more effectively balance noise suppression and reverberation suppression during the recording enhancement process, especially improving the suppression effect on late reverberation, thereby improving the speech clarity, perceived quality and overall intelligibility of the enhanced recording; at the same time, this system still maintains good computational efficiency and real-time processing capabilities, making it suitable for resource-constrained devices or real-time speech processing scenarios.

[0050] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0051] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0052] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A recording noise reduction method based on artificial intelligence, characterized in that, The method includes: A short-time Fourier transform is performed on the recording signal to be processed to obtain the observed time-frequency coefficients arranged by frequency and time units, and the current input time-frequency window and noise suppression sample vector are constructed based on the observed time-frequency coefficients. The current input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the first time-frequency context embedding representation, and the noise suppression decoding network in the artificial intelligence enhancement model outputs noise suppression filtering parameters according to the first time-frequency context embedding representation; The noise suppression sample vector is linearly filtered using the noise suppression filtering parameters to obtain the noise suppression result. A decoding input time-frequency window and a reverberation suppression sample vector are constructed based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. The decoding input time-frequency window is input into the encoding network in the artificial intelligence enhancement model to extract the second time-frequency context embedding representation. The reverberation suppression decoding network in the artificial intelligence enhancement model outputs reverberation suppression filtering parameters based on the second time-frequency context embedding representation. The reverberation suppression filtering parameters are then used to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate. The late reflection estimate is then subtracted from the noise suppression result to obtain the enhancement time-frequency coefficients. Perform an inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

2. The recording noise reduction method based on artificial intelligence according to claim 1, characterized in that, The recording signal to be processed is obtained by superimposing the reverberant speech formed by convolving the target speech with the acoustic impulse response and the background noise. The acoustic impulse response is divided into an early component containing direct sound and early reflections and a late component containing late reflections. The noise-reduced recording signal is used to retain the speech information corresponding to the early component and suppress the background noise and the reverberation information corresponding to the late component.

3. The recording noise reduction method based on artificial intelligence according to claim 1, characterized in that, The current input time-frequency window is composed of the observation time-frequency coefficients of multiple consecutive frames arranged around the current time unit and all frequency units; the noise suppression sample vector is composed of the observation time-frequency coefficients of the current frequency unit corresponding to the current time unit and the observation time-frequency coefficients of its adjacent time series; the reverberation suppression sample vector is composed of the continuous historical time-frequency coefficients of the current frequency unit corresponding to the time-frequency signal used for reverberation suppression calculation, after a delay relative to the current time unit; the vector length of the reverberation suppression sample vector is greater than the vector length of the noise suppression sample vector, so that the reverberation suppression decoding network represents a time correlation longer than that of noise suppression.

4. The recording noise reduction method based on artificial intelligence according to claim 1, characterized in that, The artificial intelligence enhancement model further includes a masking decoding network, and the encoding network is a shared encoding network; the method further includes dividing the observed time-frequency coefficients into low-frequency intervals and high-frequency intervals according to preset boundary frequency units, extracting complex time-frequency features from the low-frequency intervals, extracting equivalent rectangular bandwidth features obtained by amplitude spectrum downsampling from the high-frequency intervals, and inputting the complex time-frequency features and the equivalent rectangular bandwidth features into the shared encoding network so that the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network can output corresponding enhancement parameters respectively.

5. The recording noise reduction method based on artificial intelligence according to claim 4, characterized in that, The masking decoding network outputs real-valued masking parameters for the high-frequency range. These real-valued masking parameters are predicted in the equivalent rectangular bandwidth domain and then interpolated back to the short-time Fourier transform domain, where they are applied to the corresponding high-frequency observation time-frequency coefficients. The noise suppression decoding network outputs complex linear filtering parameters for the low-frequency range. The reverberation suppression decoding network outputs reverberation suppression filtering parameters for the low-frequency range, where the reverberation suppression filtering parameters are complex linear prediction parameters. The enhancement result for the low-frequency range is obtained by subtracting the noise suppression result formed by the complex linear filtering parameters applied to the noise suppression sample vector from the late reflection estimate formed by the complex linear prediction parameters applied to the reverberation suppression sample vector. The enhancement result for the high-frequency range is concatenated with the enhancement result for the low-frequency range to form the enhanced time-frequency coefficients.

6. The recording noise reduction method based on artificial intelligence according to claim 5, characterized in that, The shared coding network generates an embedded representation covering all frequency units in a single forward computation. The noise suppression decoding network and the reverberation suppression decoding network output corresponding complex enhancement parameters according to the frequency units. The noise suppression sample vector and the reverberation suppression sample vector are both composed of continuous time-series time-frequency coefficients on the same frequency unit, so as to complete the enhancement operation in the low-frequency range without introducing cross-band filtering.

7. The recording noise reduction method based on artificial intelligence according to claim 4, characterized in that, The time-frequency signal used for reverberation suppression calculation is an intermediate time-frequency signal formed from the first-stage enhancement result. The method employs a two-step inference approach to perform enhancement processing. In the first step, the observed time-frequency coefficients are enhanced using the shared coding network, the masking decoding network, and the noise suppression decoding network to obtain the intermediate time-frequency signal. In the second step, the decoding input time-frequency window and the reverberation suppression sample vector are constructed based on the intermediate time-frequency signal. The reverberation suppression filter parameters are output using the shared coding network and the reverberation suppression decoding network. The reverberation suppression filter parameters are then applied to the reverberation suppression sample vector to obtain an intermediate late reflection estimate. The intermediate late reflection estimate is subtracted from the intermediate time-frequency signal to obtain the final enhanced time-frequency coefficients. In the second step, the masking decoding network and the noise suppression decoding network are not enabled. The second step performs reverberation suppression for the low-frequency range, while maintaining the first-stage enhancement result for the high-frequency range.

8. The recording noise reduction method based on artificial intelligence according to claim 4, characterized in that, The AI-enhanced model is trained as follows: a training set is constructed based on pairs of clean speech samples and noisy reverberant speech samples; the shared encoding network, the masking decoding network, and the noise suppression decoding network are initialized using pre-trained enhancement model parameters; in the early stages of training, the parameters of the remaining networks except the reverberation suppression decoding network are fixed, and only the parameters of the reverberation suppression decoding network are updated; after the reverberation suppression decoding network completes initial convergence, the shared encoding network, the masking decoding network, the noise suppression decoding network, and the reverberation suppression decoding network are jointly trained end-to-end; the joint training adopts an objective function that includes spectral consistency loss and multi-resolution loss, wherein the spectral consistency loss calculates the amplitude difference after perceptually compressing the spectral amplitude of clean speech and enhanced speech, and constrains the phase consistency; the multi-resolution loss constrains the compressed spectral amplitude difference and phase consistency at multiple analysis window scales respectively, so as to improve the perceptual quality and reverberation suppression capability of the noise-reduced recording signal.

9. The recording noise reduction method based on artificial intelligence according to claim 1, characterized in that, The time-frequency signal used for reverberation suppression calculation is the time-frequency signal corresponding to the observed time-frequency coefficients. The decoding input time-frequency window is constructed by delaying the observed time-frequency coefficients in time. The noise suppression decoding network outputs the noise suppression filtering parameters according to the first time-frequency context embedding representation extracted from the current input time-frequency window. The reverberation suppression decoding network outputs the reverberation suppression filtering parameters according to the second time-frequency context embedding representation extracted from the decoding input time-frequency window. The noise suppression result and the late reflection estimate are calculated in the same forward inference, so as to simultaneously complete background noise suppression and late reverberation suppression in a single enhancement calculation.

10. An artificial intelligence-based audio recording noise reduction system, characterized in that, include: The time-frequency transformation and input construction module is used to perform short-time Fourier transform on the recording signal to be processed, obtain the observed time-frequency coefficients arranged by frequency unit and time unit, and construct the current input time-frequency window and noise suppression sample vector based on the observed time-frequency coefficients; The noise suppression parameter estimation module is used to input the current input time-frequency window into the encoding network of the artificial intelligence enhancement model, extract the first time-frequency context embedding representation, and use the noise suppression decoding network in the artificial intelligence enhancement model to output noise suppression filtering parameters according to the first time-frequency context embedding representation; The noise suppression execution and reverberation input construction module is used to perform linear filtering on the noise suppression sample vector using the noise suppression filtering parameters to obtain the noise suppression result, and to construct a decoding input time-frequency window and a reverberation suppression sample vector based on the time-frequency signal used for reverberation suppression calculation. The reverberation suppression sample vector is introduced with a preset delay relative to the current time unit to characterize the long-term dependency information corresponding to the late reflection. The reverberation suppression parameter estimation and late reflection cancellation module is used to input the decoded input time-frequency window into the encoding network in the artificial intelligence enhancement model, extract the second time-frequency context embedding representation, and use the reverberation suppression decoding network in the artificial intelligence enhancement model to output reverberation suppression filtering parameters according to the second time-frequency context embedding representation. The module then uses the reverberation suppression filtering parameters to perform linear prediction on the reverberation suppression sample vector to obtain the late reflection estimate, and subtracts the late reflection estimate from the noise suppression result to obtain the enhancement time-frequency coefficients. The time-domain reconstruction output module is used to perform an inverse short-time Fourier transform on the enhanced time-frequency coefficients to obtain a noise-reduced recording signal.

Images