Intelligent noise reduction methods and systems
By constructing a cross-frequency domain transfer learning network and domain adversarial training, adaptive transfer and fusion of voiceprint features are achieved, solving the problem of poor performance of traditional sound noise reduction methods in complex noise environments and improving noise suppression capability and sound restoration quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SKY WING HK ELECTRONIC CO LTD
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional sound noise reduction methods are ineffective in complex noise environments and lack cross-frequency domain feature transfer mechanisms, resulting in insufficient noise suppression capabilities and loss of original sound information.
By collecting voiceprint signals and representing them in time and frequency, a cross-frequency domain transfer learning network is constructed to achieve adaptive transfer and fusion of voiceprint features. The feature transfer strategy is optimized by using domain adversarial training, and reconstructed vectors and modulation vectors are generated for dynamic noise reduction.
It improves noise suppression capability and noise reduction stability in complex environments, enhances adaptability to different noise types, and significantly improves the accuracy of the noise reduction process and the quality of sound reproduction.
Smart Images
Figure CN122135731A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to audio signal processing technology, and more particularly to intelligent noise reduction processing methods and systems. Background Technology
[0002] Voiceprint recognition and noise reduction technologies have wide applications in voice communication, audio processing, and intelligent audio systems. With the development of artificial intelligence, utilizing voiceprint features for identity recognition and voice enhancement has become a research hotspot. Traditional noise reduction methods are mainly based on signal processing theory, using techniques such as spectral subtraction and Wiener filtering to suppress noise, but their performance is limited in complex noisy environments. In recent years, deep learning technology has shown significant advantages in the field of sound processing, making voiceprint analysis and noise reduction based on deep neural networks possible, enabling more effective extraction of sound features and high-quality sound reconstruction.
[0003] Traditional audio denoising algorithms typically employ feature representations in a single frequency domain or at a single scale, failing to fully capture the rich characteristics of sound signals across different frequency domains. This results in poor denoising performance in environments with complex or rapidly changing noise levels. Existing speaker feature extraction methods lack effective cross-frequency domain feature transfer mechanisms, making it difficult to achieve adaptive fusion of speaker features across different frequency domains. Consequently, the system's adaptability to changes in sound characteristics is weak. Current audio denoising systems generally lack the ability to finely control the reconstruction process, failing to dynamically adjust denoising parameters based on specific speaker features. Therefore, denoising often results in the loss of crucial information from the original sound while eliminating noise. Summary of the Invention
[0004] The present invention provides an intelligent noise reduction processing method and system that can solve the problems in the prior art.
[0005] A first aspect of the present invention provides an intelligent noise reduction processing method, comprising: Acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals; The multiple frequency domain signals are transformed to obtain multi-scale feature vectors, and a cross-frequency domain transfer learning network is constructed. The multi-scale feature vectors are input into the cross-frequency domain transfer learning network to obtain fused feature vectors. The cross-frequency domain transfer learning network constructs feature mapping relationships between frequency domains through domain adversarial training to achieve adaptive transfer and fusion of voiceprint features. A feature library is constructed using the fused feature vectors, and the similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity. Noise reduction parameters are generated based on the feature similarity, and the noise reduction parameters include a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; A noise reduction result is generated based on the adjusted voiceprint signal to be identified, and the noise reduction result is fed back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
[0006] The voiceprint signal is converted into a time-frequency representation, and the time-frequency representation is divided into frequency domains to obtain multiple frequency domain signals, including: The voiceprint signal is preprocessed to eliminate noise interference, and the voiceprint signal is converted from time to frequency using short-time Fourier transform to obtain a time-frequency representation result. The short-time Fourier transform achieves multi-resolution analysis of the signal by adaptively adjusting the time window length, wherein the time window length is dynamically determined based on the local statistical characteristics of the signal. The energy distribution characteristics of the time-frequency representation result are calculated to construct an energy density matrix. Based on the energy density matrix, an iterative optimization algorithm is used to determine the frequency domain partitioning threshold. The iterative optimization algorithm combines the information entropy criterion and the energy concentration criterion to dynamically adjust the partitioning boundary, thereby achieving adaptive segmentation of the frequency domain. The time-frequency representation result is segmented according to the determined frequency domain partitioning threshold to obtain multiple frequency domain signals.
[0007] The cross-frequency domain transfer learning network constructs feature mapping relationships between frequency domains through domain adversarial training, achieving adaptive transfer and fusion of voiceprint features, including: The cross-frequency domain transfer learning network includes a shared feature extraction subnetwork, a domain classification subnetwork, and a task prediction subnetwork. The shared feature extraction subnetwork receives the multi-scale feature vector and outputs domain-invariant features. A gradient inversion layer is set between the shared feature extraction subnetwork and the domain classification subnetwork, so that the gradient signal generated by the domain classification subnetwork during the optimization process is inverted by the gradient inversion layer and then transmitted to the shared feature extraction subnetwork, driving the shared feature extraction subnetwork to learn feature representations that eliminate frequency domain differences; The domain-invariant features are simultaneously input into the domain classification subnetwork and the task prediction subnetwork. By minimizing the classification loss of the task prediction subnetwork and maximizing the domain confusion loss of the domain classification subnetwork, a feature mapping relationship is established that makes the signals of different frequency domains tend to be distributed and aligned in the feature space, thereby realizing the adaptive transfer and fusion of voiceprint features.
[0008] A feature library is constructed using the fused feature vectors. The similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity, which includes: A multidimensional feature library is constructed using the fused feature vectors. The multidimensional feature library organizes the fused feature vectors through a hierarchical index structure, wherein the hierarchical index structure divides the fused feature vectors into multiple feature clusters according to the degree of feature similarity. Hash functions are calculated for each of the multiple feature clusters based on the hierarchical index structure. The voiceprint signal to be identified is converted into a retrieval feature vector, and candidate feature clusters are determined through the hierarchical index structure. The hash function corresponding to the candidate feature cluster is invoked to map the retrieved feature vector to the multidimensional feature library, and the feature similarity is obtained by calculating the vector distance in the hash space.
[0009] Noise reduction parameters are generated based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the speaker signal are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector, including: Noise reduction parameters are generated based on the feature similarity. The noise reduction parameters include a reconstruction vector and a modulation vector. The reconstruction vector and the modulation vector are mapped to the frequency domain and the amplitude domain, respectively, to form a dual-domain noise reduction strategy. The speaker signal is processed based on the dual-domain noise reduction strategy. The frequency features of the speaker signal are reconstructed using a reconstruction vector mapped to the frequency domain. The optimal reconstruction coefficient is calculated iteratively, and the frequency components are adjusted according to the optimal reconstruction coefficient. The amplitude features of the reconstructed signal are dynamically adjusted using a modulation vector mapped to the amplitude domain.
[0010] Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated, and the noise reduction and enhancement result is fed back to the cross-frequency domain transfer learning network. The feature transfer strategy is optimized through the domain adversarial training method, including: Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated. The noise reduction and enhancement result is quantitatively evaluated by calculating the signal-to-noise ratio and spectral distortion, and optimization feedback information is generated. The optimized feedback information is fed back to the cross-frequency domain transfer learning network, and the network parameters are updated through the domain adversarial training method. The network parameter update process adopts an adaptive learning rate strategy, and the adaptive learning rate is dynamically adjusted according to the noise reduction effect evaluation results to optimize the feature transfer strategy.
[0011] A second aspect of the present invention provides an intelligent noise reduction processing system, comprising: The acquisition module is used to acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals. The transformation module is used to transform the multiple frequency domain signals respectively, obtain multi-scale feature vectors, construct a cross-frequency domain transfer learning network, input the multi-scale feature vectors into the cross-frequency domain transfer learning network to obtain fused feature vectors, and the cross-frequency domain transfer learning network constructs the feature mapping relationship between frequency domains through domain adversarial training to realize adaptive transfer and fusion of voiceprint features. The calculation module is used to construct a feature library using the fused feature vector, calculate the similarity between the voiceprint signal to be identified and the features in the feature library, and obtain the feature similarity. An adjustment module is used to generate noise reduction parameters based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; The feedback module is used to generate a noise reduction result based on the adjusted voiceprint signal to be identified, and to feed the noise reduction result back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
[0012] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0013] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0014] The beneficial effects of this application are as follows: The intelligent noise reduction processing method provided by this invention collects voiceprint signals and converts them into time-frequency representation. It divides the frequency domain according to the energy distribution threshold, effectively solving the problem of uneven noise processing in different frequency domains by traditional noise reduction methods and improving the noise suppression capability in complex environments.
[0015] By establishing a cross-frequency domain transfer learning network to achieve adaptive transfer and fusion of voiceprint features, this method can capture multi-scale features of voiceprint signals and construct feature mapping relationships between frequency domains through domain adversarial training, thereby enhancing the system's adaptability to different noise types and improving the stability of noise reduction.
[0016] This invention generates noise reduction parameters containing reconstruction vectors and modulation vectors based on feature similarity. By reconstructing and dynamically adjusting the speaker signal, and feeding the noise reduction enhancement results back to the learning network for optimization, a closed-loop adaptive mechanism is formed, which significantly improves the accuracy of the noise reduction process and the sound restoration quality, and is suitable for various complex acoustic environments. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the intelligent noise reduction processing method according to an embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.
[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0020] Figure 1 This is a flowchart illustrating the intelligent noise reduction processing method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals; The multiple frequency domain signals are transformed to obtain multi-scale feature vectors, and a cross-frequency domain transfer learning network is constructed. The multi-scale feature vectors are input into the cross-frequency domain transfer learning network to obtain fused feature vectors. The cross-frequency domain transfer learning network constructs feature mapping relationships between frequency domains through domain adversarial training to achieve adaptive transfer and fusion of voiceprint features. A feature library is constructed using the fused feature vectors, and the similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity. Noise reduction parameters are generated based on the feature similarity, and the noise reduction parameters include a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; A noise reduction result is generated based on the adjusted voiceprint signal to be identified, and the noise reduction result is fed back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
[0021] In one optional implementation, converting the voiceprint signal into a time-frequency representation, and dividing the time-frequency representation into frequency domains to obtain multiple frequency domain signals includes: The voiceprint signal is preprocessed to eliminate noise interference, and the voiceprint signal is converted from time to frequency using short-time Fourier transform to obtain a time-frequency representation result. The short-time Fourier transform achieves multi-resolution analysis of the signal by adaptively adjusting the time window length, wherein the time window length is dynamically determined based on the local statistical characteristics of the signal. The energy distribution characteristics of the time-frequency representation result are calculated to construct an energy density matrix. Based on the energy density matrix, an iterative optimization algorithm is used to determine the frequency domain partitioning threshold. The iterative optimization algorithm combines the information entropy criterion and the energy concentration criterion to dynamically adjust the partitioning boundary, thereby achieving adaptive segmentation of the frequency domain. The time-frequency representation result is segmented according to the determined frequency domain partitioning threshold to obtain multiple frequency domain signals.
[0022] The input speaker signal is preprocessed to eliminate noise interference. Wiener filtering is employed, which adaptively suppresses background noise by estimating the power spectral ratio of the signal to the noise. Specifically, the input signal is first segmented into frames, each with 512 sampling points and a 50% overlap between adjacent frames. Short-time energy and zero-crossing rate are calculated for each frame to determine if it is a noise segment. If the short-time energy is below a preset threshold (e.g., 0.3 times the average signal energy) and the zero-crossing rate is above a preset threshold (e.g., 1.5 times the average zero-crossing rate), the frame is marked as a noise segment. By collecting samples from multiple noise segments, the power spectral characteristics of the noise are estimated, and Wiener filter coefficients are constructed to filter the entire signal. Experimental results show that this preprocessing method can effectively improve the signal-to-noise ratio by 5-8 dB.
[0023] After preprocessing, the audioprint signal is converted to time-frequency using a short-time Fourier transform (SFT). In this embodiment, the SFT achieves multi-resolution analysis of the signal by adaptively adjusting the time window length. The window length is dynamically determined based on the local statistical characteristics of the signal, primarily considering the local stationarity of the signal. During implementation, the signal is first divided into analysis units of 1024 points. The autocorrelation function is calculated for each analysis unit, and the decay rate of the autocorrelation function is detected. If the decay rate is fast, it indicates rapid local changes in the signal, and a shorter window should be selected to improve time resolution; if the decay rate is slow, it indicates relatively stable local conditions in the signal, and a longer window should be selected to improve frequency resolution.
[0024] In practical implementation, the window length set is defined as {256, 512, 1024, 2048}, and the window is selected based on the number of delay points required for the autocorrelation function to decay to 50% of its initial value. For example, if the number of points required to decay to 50% is less than 64, a 256-point window is selected; if it is between 64 and 128, a 512-point window is selected; if it is between 128 and 256, a 1024-point window is selected; and if it is greater than 256, a 2048-point window is selected. After selecting the window length for each analysis unit, the Hanning window function is applied for windowing, and then a Fast Fourier Transform is performed. In experiments, this adaptive windowing method can improve the clarity of time-frequency representation by an average of 15% compared to the fixed window method, especially when processing speech segments containing a mixture of consonants and vowels.
[0025] After performing a short-time Fourier transform, a time-frequency domain representation matrix S is obtained, with dimensions equal to the number of time frames multiplied by the number of frequency points. Next, the energy distribution characteristics of the time-frequency transformation result are calculated to construct an energy density matrix. For each element S(t, f) in matrix S, its energy value |S(t, f)|² is calculated, resulting in the energy density matrix E. To highlight frequency bands with concentrated energy, the cumulative energy distribution along the frequency dimension of the energy density matrix E is calculated. Specifically, for each frequency point f, the proportion of cumulative energy from 0 Hz to f to the total energy is calculated, forming a cumulative energy distribution curve C(f).
[0026] This implementation uses an iterative optimization algorithm to determine the frequency domain partitioning threshold based on the energy density matrix design for non-uniform frequency domain partitioning. The algorithm dynamically adjusts the partitioning boundary by combining the information entropy criterion and the energy concentration criterion. Initially, the frequency domain is uniformly divided into K sub-bands (e.g., K=8). For each sub-band, its information entropy H and energy concentration G are calculated. The information entropy H reflects the complexity of the energy distribution within the sub-band, and the energy concentration G represents the degree of energy concentration within the sub-band. If the information entropy H of a sub-band is higher than a set threshold (e.g., normalized entropy value 0.8) and the energy concentration G is lower than a set threshold (e.g., 0.3), it indicates that the energy distribution within that sub-band is complex and requires further subdivision.
[0027] During the iteration process, for each sub-band that needs to be further subdivided, the optimal splitting point is found near its boundary. Specifically, the splitting point is moved in 1% increments within a ±15% range near the boundary. For each position, a weighted sum of the information entropy and energy concentration of the two new sub-bands after the split is calculated, and the position that yields the optimal weighted sum is selected as the new splitting point. The iteration continues until the information entropy and energy concentration of all sub-bands meet the requirements, or the preset maximum number of sub-bands (e.g., 16) is reached.
[0028] In a practical test case, a voiceprint signal containing the two syllables "Hello" was processed. After initially dividing it evenly into 8 sub-bands, the iterative optimization algorithm finally determined 12 non-uniform sub-bands. These sub-bands have a finer division in the frequency concentration regions of the vowel segments (such as 300 - 800 Hz and 1500 - 2500 Hz), while the division is coarser in the high-frequency regions with weaker energy (such as 4000 - 8000 Hz). Compared with the uniform division, this non-uniform division method reduces the error rate from 6.5% to 4.2% in the subsequent voiceprint recognition task.
[0029] According to the determined frequency-domain division threshold, the time-frequency representation form is segmented to obtain multiple frequency-domain signals. For each divided frequency band, the trajectory of its energy changing with time is extracted to form a frequency-domain feature vector. These feature vectors capture the energy change patterns of different frequency bands and provide rich discriminant information for voiceprint recognition. The extraction of the frequency-domain signal uses the inverse short-time Fourier transform. The energy outside the target frequency band is set to zero in the time-frequency domain, and then the time-domain signal is reconstructed through the inverse transform to obtain the sub-signals of each frequency band.
[0030] In this embodiment, through adaptive time-frequency analysis and dynamic frequency-domain division, the fine processing of the voiceprint signal is realized, laying a solid foundation for subsequent voiceprint feature extraction and recognition.
[0031] In an optional embodiment, the cross-frequency-domain transfer learning network constructs a feature mapping relationship between frequency domains through domain adversarial training to achieve the adaptive transfer and fusion of voiceprint features, including: The cross-frequency-domain transfer learning network includes a shared feature extraction sub-network, a domain classification sub-network, and a task prediction sub-network. The shared feature extraction sub-network receives the multi-scale feature vector and outputs domain-invariant features; A gradient reversal layer is set between the shared feature extraction sub-network and the domain classification sub-network, so that the gradient signal generated during the optimization of the domain classification sub-network is passed to the shared feature extraction sub-network after the sign is reversed by the gradient reversal layer, driving the shared feature extraction sub-network to learn to eliminate the feature representation of frequency-domain differences; The domain-invariant features are simultaneously input into the domain classification sub-network and the task prediction sub-network. By minimizing the classification loss of the task prediction sub-network and simultaneously maximizing the domain confusion loss of the domain classification sub-network, a feature mapping relationship that makes different frequency-domain signals tend to be distributed in alignment in the feature space is established to achieve the adaptive transfer and fusion of voiceprint features.
[0032] Multiple frequency domain signals are transformed separately to obtain multi-scale feature vectors. Each frequency domain signal undergoes multi-level decomposition and transformation, extracting local time-frequency feature components at different scale levels. For example, wavelet transform is used to perform a three-level decomposition of the frequency domain signal, obtaining low-frequency approximation components and high-frequency detail components. For an original frequency domain signal of length 1024, the first level of decomposition yields a low-frequency approximation component and a high-frequency detail component of length 512; the second level yields a lower-frequency approximation component and a mid-frequency detail component of length 256; and the third level yields a lowest-frequency approximation component and a mid-low-frequency detail component of length 128.
[0033] These components at different levels represent the characteristics of the signal in different frequency ranges, enabling the capture of the signal's multi-scale features. These local time-frequency feature components are concatenated in a progressive scale order to form a multi-scale feature vector. In this embodiment, the lowest frequency approximation component (128) of the third layer, the mid-to-low frequency detail component (128), the mid-frequency detail component (256) of the second layer, and the high-frequency detail component (512) of the first layer are concatenated in order from low to high frequency to form a multi-scale feature vector of length 1024.
[0034] A cross-frequency domain transfer learning network is constructed, comprising a shared feature extraction subnetwork, a domain classification subnetwork, and a task prediction subnetwork. The shared feature extraction subnetwork consists of three fully connected layers, each containing 256, 128, and 64 neurons respectively, using ReLU as the activation function to map multi-scale feature vectors to a 64-dimensional domain-invariant feature space. The domain classification subnetwork consists of two fully connected layers, each containing 32 neurons, with the final layer using the Softmax activation function to predict which frequency domain the features originate from. The task prediction subnetwork also consists of two fully connected layers, each containing 32 neurons, with the final layer using the Softmax activation function to predict the target task.
[0035] A gradient inversion layer is placed between the shared feature extraction subnetwork and the domain classification subnetwork. This layer does not change the input data during forward propagation, but multiplies the gradient by -1 during backpropagation, thus inverting the gradient sign. For example, when the loss function value of the domain classification subnetwork is 0.8 and the gradient calculated during backpropagation is 0.3, after passing through the gradient inversion layer, the gradient passed to the shared feature extraction subnetwork becomes -0.3. This mechanism drives the shared feature extraction subnetwork to learn feature representations that eliminate frequency domain differences, making it difficult to distinguish samples from different frequency domains in the feature space.
[0036] During network training, domain-invariant features are simultaneously input into both the domain classification sub-network and the task prediction sub-network. The task prediction sub-network uses the cross-entropy loss function, calculating the loss between the prediction result and the true label for each sample, and updating the network parameters through backpropagation to make the prediction results more accurate. The domain classification sub-network also uses the cross-entropy loss function, but due to the presence of the gradient inversion layer, the shared feature extraction sub-network is actually trying to maximize the loss of the domain classification sub-network, i.e., increasing the difficulty of domain classification. By simultaneously optimizing these two objectives, the network establishes a feature mapping relationship that makes the signals of different frequency domains tend to be aligned in the feature space.
[0037] In a specific implementation case, this method is applied to a cross-frequency domain radar signal recognition task. Radar signals from three different frequency bands (S-band: 2-4 GHz, C-band: 4-8 GHz, X-band: 8-12 GHz) were collected, each band containing eight different types of radar signals, with 500 samples for each type. The S-band data was used as the source domain, the X-band data as the target domain, and the C-band data was used to verify the model's generalization ability. A three-layer wavelet decomposition was performed on each frequency domain signal to obtain multi-scale feature vectors. The constructed cross-frequency domain transfer learning network includes the above three sub-network structures, where the output dimension of the domain classification sub-network is 3 (corresponding to the three frequency bands), and the output dimension of the task prediction sub-network is 8 (corresponding to the eight radar signal types). During the training phase, domain adversarial training was performed using labeled S-band data and unlabeled X-band data for a total of 100 training rounds, with a batch size of 64 and a learning rate of 0.001. After training, the model achieved a recognition accuracy of 92.3% on the X-band test set, which is 26.6 percentage points higher than the baseline model (65.7%) without cross-frequency domain transfer learning. It also achieved an accuracy of 88.5% on the C-band test set, demonstrating the effectiveness and generalization ability of the proposed method in cross-frequency domain signal processing.
[0038] Through the above implementation methods, this invention utilizes multi-scale feature extraction and domain adversarial training techniques to successfully achieve knowledge transfer between different frequency domains, solving the problem of model performance degradation caused by frequency domain differences, and providing an effective technical solution for the field of multi-frequency domain signal processing.
[0039] In one optional implementation, a feature library is constructed using the fused feature vector, and the similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity, including: A multidimensional feature library is constructed using the fused feature vectors. The multidimensional feature library organizes the fused feature vectors through a hierarchical index structure, wherein the hierarchical index structure divides the fused feature vectors into multiple feature clusters according to the degree of feature similarity. Hash functions are calculated for each of the multiple feature clusters based on the hierarchical index structure. The voiceprint signal to be identified is converted into a retrieval feature vector, and candidate feature clusters are determined through the hierarchical index structure. The hash function corresponding to the candidate feature cluster is invoked to map the retrieved feature vector to the multidimensional feature library, and the feature similarity is obtained by calculating the vector distance in the hash space.
[0040] A multi-dimensional feature library is constructed using fused feature vectors, organized with a hierarchical index structure. A large number of collected voiceprint samples are processed using feature extraction algorithms to obtain corresponding fused feature vectors. Each vector typically contains 256 to 1024 dimensions, representing the audio characteristics of the voiceprint. Based on the similarity of these feature vectors, the system divides them into multiple feature clusters using a clustering method. For example, the fused feature vectors of 10,000 voiceprint samples are divided into 100 feature clusters, each containing voiceprint feature vectors with high similarity.
[0041] In constructing the hierarchical index structure, a top-down approach is adopted for partitioning. All fused feature vectors are treated as a large feature cluster, and then, based on the Euclidean distance between the feature vectors, the K-means clustering algorithm is used to divide it into several sub-clusters. For each sub-cluster, the system continues to recursively apply the same partitioning method until the size of each feature cluster is less than a preset threshold (e.g., 100 samples) or reaches a predetermined hierarchical depth (e.g., 4 layers). This hierarchical structure makes similar voiceprint features spatially closer, facilitating subsequent rapid retrieval.
[0042] For each feature cluster, a corresponding hash function is calculated to map the high-dimensional feature vector to a low-dimensional hash space. In this embodiment, the system generates eight locality-sensitive hash functions for each feature cluster, each defined by a random hyperplane. Specifically, the system randomly generates a vector with the same dimension as the feature vector as the normal vector, and then calculates the dot product between the feature vector to be processed and the normal vector. If the dot product result is greater than zero, the corresponding hash bit is 1; otherwise, it is 0. Through these eight hash functions, the system maps each feature vector to an 8-bit hash code.
[0043] In the voiceprint recognition process, the voiceprint signal to be identified undergoes preprocessing, including noise reduction, framing, and window function processing. For example, for a 16kHz sampling rate audio signal, the system divides it into 25-millisecond frames with a 10-millisecond frame shift and applies a Hamming window function to reduce spectral leakage. Subsequently, the system extracts the acoustic features of each frame, such as Mel-frequency cepstral coefficients, fundamental frequency, and spectral flux, and converts these raw features into a fixed-dimensional voiceprint feature vector using a deep neural network model.
[0044] The extracted voiceprint feature vectors are mapped to hash codes using the same hash function. Due to the hierarchical index structure, the system can quickly locate the feature cluster to which the feature vector to be retrieved belongs. Specifically, starting from the top layer of the hierarchical structure, the system calculates the distance between the vector to be retrieved and the center of each sub-cluster, selects the nearest sub-cluster, and continues searching downwards until the leaf node layer is reached.
[0045] After determining the target feature cluster, the system uses the hash function corresponding to the feature cluster to convert the voiceprint feature vector to be identified into a hash code, and then performs hash matching within that feature cluster. The system quickly filters out potential matching candidates by calculating the Hamming distance between the hash code of the voiceprint to be identified and the hash codes of samples in the feature library. For example, if the Hamming distance between two hash codes is less than 2 (i.e., at most 2 bits are different), it is considered a potential match.
[0046] For the initially selected candidates, the precise distance in the original feature vector space is calculated. In this embodiment, the system uses cosine similarity as the distance metric, that is, the cosine value of the angle between two vectors is calculated. The cosine similarity range is [-1, 1], and the larger the value, the more similar the two vectors are. The system considers matches with a cosine similarity greater than a preset threshold (e.g., 0.85) as valid recognition results.
[0047] To illustrate with a concrete example, suppose the feature database stores the voiceprint features of 1000 registered users, with 3-5 voiceprint samples for each user. The system organizes the fused feature vectors of these samples into 20 feature clusters using a hierarchical index structure. When a new voiceprint needs to be identified, the system first locates the feature cluster through hierarchical search, which contains 50 samples. The system uses the hash function of this cluster to convert the voiceprint to be identified into a hash code and calculates the Hamming distance with samples within the cluster, selecting 5 candidate voiceprints with a Hamming distance less than 3. Finally, the system calculates the cosine similarity between the voiceprint to be identified and the fused feature vectors of these 5 candidate voiceprints, obtaining a matching result with a maximum similarity of 0.92, exceeding the threshold of 0.85. Therefore, the system successfully identifies the user corresponding to this voiceprint.
[0048] Through the above implementation methods, a highly efficient voiceprint recognition process is achieved. In particular, in large-scale voiceprint databases, the hierarchical index structure and hash mapping technology significantly improve retrieval efficiency, enabling the system to complete voiceprint matching within milliseconds while ensuring a high recognition accuracy.
[0049] In one optional implementation, noise reduction parameters are generated based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the speaker signal are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector, including: Noise reduction parameters are generated based on the feature similarity. The noise reduction parameters include a reconstruction vector and a modulation vector. The reconstruction vector and the modulation vector are mapped to the frequency domain and the amplitude domain, respectively, to form a dual-domain noise reduction strategy. The speaker signal is processed based on the dual-domain noise reduction strategy. The frequency features of the speaker signal are reconstructed using a reconstruction vector mapped to the frequency domain. The optimal reconstruction coefficient is calculated iteratively, and the frequency components are adjusted according to the optimal reconstruction coefficient. The amplitude features of the reconstructed signal are dynamically adjusted using a modulation vector mapped to the amplitude domain.
[0050] In the feature similarity calculation stage, the input voiceprint signal is processed by framing, with each frame length set to 25 milliseconds and a frame shift of 10 milliseconds to ensure sufficient overlap between frames to maintain signal continuity. A short-time Fourier transform is applied to each frame to obtain time-frequency domain features, extracting characteristic parameters including spectral energy distribution, spectral centroid, and harmonic structure. Based on these characteristic parameters, feature vectors are constructed, and the cosine similarity between adjacent frames is calculated. For example, when the feature vectors of frame i and frame i-1 are Vi and Vi-1 respectively, the calculated similarity is 0.85, indicating a high similarity between the two frames. For all frames, a feature similarity matrix is established, which contains correlation information for each time-frequency point of the voiceprint signal.
[0051] Based on the feature similarity matrix, denoising parameters are generated, including a reconstruction vector and a modulation vector. The reconstruction vector is mainly used for signal optimization in the frequency domain, while the modulation vector operates in the amplitude domain. The reconstruction vector is generated based on a weighted average of feature similarity. Frames with similarity above a threshold of 0.7 are assigned higher weights, such as 0.9; frames with similarity below the threshold are assigned lower weights, such as 0.3. This weighting mechanism constructs a reconstruction vector R for frequency features, with dimensions matching the number of frequency components. The modulation vector is generated considering the signal's energy distribution and noise estimation. For frequencies with low signal-to-noise ratios, the corresponding components of the modulation vector have smaller values, such as 0.2; for frequencies with high signal-to-noise ratios, the corresponding components have larger values, such as 0.9. This constructed modulation vector M is used for subsequent dynamic adjustment of the amplitude features.
[0052] In the dual-domain noise reduction strategy, the reconstructed vector R is mapped to the frequency domain, and the modulation vector M is mapped to the amplitude domain. Frequency domain mapping is achieved by multiplying the reconstructed vector R by the spectral components of the original signal, with each frequency component adjusted according to its corresponding weight in the reconstructed vector. For example, for the original spectral component F(k) and the reconstructed vector component R(k), the reconstructed frequency feature is F'(k) = F(k) × R(k). If the original component at a certain frequency point k is 10, and the corresponding value of the reconstructed vector is 0.8, then the reconstructed component is 8. This process achieves the reconstruction mapping of frequency features.
[0053] To achieve optimal reconstruction results, an iterative method is used to calculate the optimal reconstruction coefficients. The initial reconstruction coefficients are set to 0.5. Then, based on the difference between the reconstructed result and the target signal, the built-in reconstruction parameter table in the system is gradually adjusted. The reconstruction parameter table contains adjustment coefficients for different frequency bands, such as 0.75 for the low-frequency band (20Hz-200Hz), 0.85 for the mid-frequency band (200Hz-2000Hz), and 0.65 for the high-frequency band (2000Hz-8000Hz). During the iteration process, if the signal-to-noise ratio (SNR) improvement of the current reconstruction result is less than 2dB, the reconstruction coefficients are adjusted further; if the SNR improvement is less than 0.5dB after three consecutive iterations, the optimal state is considered reached, and the iteration stops. The finally determined optimal reconstruction coefficients are used to adjust the frequency components and improve the frequency characteristic quality of the acoustic signature signal.
[0054] Dynamic adjustment of the amplitude characteristic is achieved through the modulation vector M. The reconstructed spectral signal is converted into an amplitude representation and weighted with the modulation vector M. For the amplitude A(k) and the modulation vector component M(k), the adjusted amplitude characteristic is A'(k) = A(k) × M(k). This dynamic adjustment adaptively adjusts the amplitude based on the signal-to-noise ratio (SNR) at different frequency points. For example, when the estimated SNR at a certain frequency point is 5dB, the corresponding modulation vector component is 0.7, and the original amplitude of 10 is adjusted to 7; when the estimated SNR is 15dB, the modulation vector component is 0.95, and the original amplitude is adjusted to 9.5. This mechanism ensures effective suppression of noise interference while preserving valid acoustic signature information.
[0055] Through the above dual-domain noise reduction processing, the voiceprint signal was effectively enhanced. In actual testing, the average signal-to-noise ratio of the voiceprint signal before processing was 8dB, which was increased to 15dB after processing, and the voiceprint recognition accuracy increased from 78% to 94%. This method is not only suitable for mild background noise scenarios in indoor environments, but can also handle moderate noise pollution in complex environments, demonstrating strong adaptability and robustness.
[0056] In practical applications, a 10-second input speech signal with a sampling rate of 16kHz is preprocessed and then processed using the method described above. The processing time is approximately 150 milliseconds, meeting real-time processing requirements. By appropriately setting the update frequencies of the reconstruction vector and modulation vector, computational complexity and noise reduction effectiveness can be balanced, adapting to the performance limitations of different hardware platforms.
[0057] In one optional implementation, a noise reduction and enhancement result is generated based on the adjusted speaker signal, and the noise reduction and enhancement result is fed back to the cross-frequency domain transfer learning network. The feature transfer strategy is optimized through the domain adversarial training method, including: Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated. The noise reduction and enhancement result is quantitatively evaluated by calculating the signal-to-noise ratio and spectral distortion, and optimization feedback information is generated. The optimized feedback information is fed back to the cross-frequency domain transfer learning network, and the network parameters are updated through the domain adversarial training method. The network parameter update process adopts an adaptive learning rate strategy, and the adaptive learning rate is dynamically adjusted according to the noise reduction effect evaluation results to optimize the feature transfer strategy.
[0058] The system receives the regulated voiceprint signal, which refers to the signal obtained after preprocessing the original voiceprint signal through a preprocessing module. Preprocessing may include, but is not limited to, framing, windowing, and spectral analysis. For example, the system can frame the original voiceprint signal according to a 20-millisecond frame length and a 10-millisecond frame shift, then apply a Hamming window to each frame, and finally convert the time-domain signal to a frequency-domain signal using a Fast Fourier Transform.
[0059] Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated through a cross-frequency domain transfer learning network. This network consists of an encoder and a decoder. The encoder is responsible for extracting key features from the voiceprint signal, and the decoder is responsible for reconstructing these features into a clear audio signal. In a specific example, the encoder contains four convolutional layers, each with a kernel size of 3×3 and channel numbers of 32, 64, 128, and 256, respectively. Each convolutional layer is followed by a ReLU activation function and a max-pooling layer. The decoder contains four deconvolutional layers, with a structure symmetrical to the encoder, to restore the original resolution of the signal.
[0060] After the noise reduction and enhancement results are generated, they are quantitatively evaluated by calculating the signal-to-noise ratio (SNR) and spectral distortion. The SNR calculation uses a segmented weighting method, assigning different weights to signals in different frequency bands to reflect the differences in human ear sensitivity to different frequencies. For example, mid-frequency signals in the 1000Hz-4000Hz range are given a higher weight of 0.5, while frequency bands below 250Hz and above 8000Hz are given a lower weight of 0.1. In practical applications, for a speech signal containing environmental noise, the SNR before noise reduction is 5.3dB, which is improved to 15.8dB after system processing.
[0061] Spectral distortion is calculated by comparing the differences in spectral characteristics between the original clean signal and the denoised signal. Logarithmic spectral distance is used as a metric for spectral distortion, calculating the Euclidean distance between the logarithmic amplitude spectra of the two signals at various frequency points. In a practical case, for a speech signal in a moderately noisy environment, the spectral distortion before denoising was 1.89, which was reduced to 0.43 after denoising, indicating that the system effectively preserved the main spectral features of the speech.
[0062] Based on the calculation results of signal-to-noise ratio (SNR) and spectral distortion, a comprehensive score is generated as optimization feedback information. This comprehensive score adopts a weighted average method, with the improvement in SNR accounting for 60% of the weight and the improvement in spectral distortion accounting for 40%. For example, when the SNR is improved by 10dB and the spectral distortion is improved by 1.5, the comprehensive score calculated by the system is 7.8 points (out of 10).
[0063] The optimized feedback is fed back to the cross-frequency domain transfer learning network, and the network parameters are updated through domain adversarial training. Domain adversarial training consists of two main components: a feature extractor and a domain discriminator. The feature extractor attempts to extract features shared by the source and target domains, while the domain discriminator attempts to distinguish which domain the features originate from. During training, the goal of the feature extractor is to generate features that are difficult for the domain discriminator to distinguish, thereby achieving effective cross-domain feature transfer.
[0064] The network parameter update process employs an adaptive learning rate strategy, dynamically adjusting the learning rate based on the noise reduction effect evaluation results. Specifically, when the overall score does not improve significantly after three consecutive iterations (the improvement is less than 0.1 points), the system multiplies the current learning rate by 0.8 to refine the parameter search space; when the score shows a significant improvement (the improvement is greater than 0.5 points), the current learning rate is temporarily maintained to fully explore the current favorable direction.
[0065] In a real-world application, the initial learning rate was set to 0.001 and remained stable for the first 20 iterations, then dynamically adjusted based on score changes. By the 50th iteration, the learning rate had decreased to 0.00032, at which point the model performance stabilized, achieving a comprehensive score of 9.3. On a test set with diverse noise types, the system achieved an average improvement in signal-to-noise ratio of 12.5 dB and a reduction in spectral distortion of 1.75.
[0066] This feedback optimization mechanism enables the noise reduction strategy to be adaptively adjusted for different types of noise environments. For example, for steady-state background noise (such as air conditioner noise or computer fan noise), the system tends to maintain a high smoothness parameter value (about 0.85), while for sudden noise (such as knocking or door closing), the system will automatically reduce the smoothness parameter (about 0.4) to respond quickly to noise changes.
[0067] Experiments show that, compared with traditional fixed-parameter noise reduction methods, this system significantly improves speech clarity and listening comfort, with subjective scores increasing by 25.6% and user satisfaction increasing by 31.2%. In particular, the system demonstrates strong robustness in complex noisy environments (such as background noise in coffee shops and traffic noise), effectively preserving speech details while suppressing background noise interference.
[0068] A second aspect of the present invention provides an intelligent noise reduction processing system, comprising: The acquisition module is used to acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals. The transformation module is used to transform the multiple frequency domain signals respectively, obtain multi-scale feature vectors, construct a cross-frequency domain transfer learning network, input the multi-scale feature vectors into the cross-frequency domain transfer learning network to obtain fused feature vectors, and the cross-frequency domain transfer learning network constructs the feature mapping relationship between frequency domains through domain adversarial training to realize adaptive transfer and fusion of voiceprint features. The calculation module is used to construct a feature library using the fused feature vector, calculate the similarity between the voiceprint signal to be identified and the features in the feature library, and obtain the feature similarity. An adjustment module is used to generate noise reduction parameters based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; The feedback module is used to generate a noise reduction result based on the adjusted voiceprint signal to be identified, and to feed the noise reduction result back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
[0069] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0070] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0071] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.
[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intelligent noise reduction processing method, characterized in that, include: Acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals; The multiple frequency domain signals are transformed to obtain multi-scale feature vectors, and a cross-frequency domain transfer learning network is constructed. The multi-scale feature vectors are input into the cross-frequency domain transfer learning network to obtain fused feature vectors. The cross-frequency domain transfer learning network constructs feature mapping relationships between frequency domains through domain adversarial training to achieve adaptive transfer and fusion of voiceprint features. A feature library is constructed using the fused feature vectors, and the similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity. Noise reduction parameters are generated based on the feature similarity, and the noise reduction parameters include a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; A noise reduction result is generated based on the adjusted voiceprint signal to be identified, and the noise reduction result is fed back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
2. The method according to claim 1, characterized in that, The voiceprint signal is converted into a time-frequency representation, and the time-frequency representation is divided into frequency domains to obtain multiple frequency domain signals, including: The voiceprint signal is preprocessed to eliminate noise interference, and the voiceprint signal is converted from time to frequency using short-time Fourier transform to obtain a time-frequency representation result. The short-time Fourier transform achieves multi-resolution analysis of the signal by adaptively adjusting the time window length, wherein the time window length is dynamically determined based on the local statistical characteristics of the signal. The energy distribution characteristics of the time-frequency representation result are calculated to construct an energy density matrix. Based on the energy density matrix, an iterative optimization algorithm is used to determine the frequency domain partitioning threshold. The iterative optimization algorithm combines the information entropy criterion and the energy concentration criterion to dynamically adjust the partitioning boundary, thereby achieving adaptive segmentation of the frequency domain. The time-frequency representation result is segmented according to the determined frequency domain partitioning threshold to obtain multiple frequency domain signals.
3. The method according to claim 1, characterized in that, The cross-frequency domain transfer learning network constructs feature mapping relationships between frequency domains through domain adversarial training, achieving adaptive transfer and fusion of voiceprint features, including: The cross-frequency domain transfer learning network includes a shared feature extraction subnetwork, a domain classification subnetwork, and a task prediction subnetwork. The shared feature extraction subnetwork receives the multi-scale feature vector and outputs domain-invariant features. A gradient inversion layer is set between the shared feature extraction subnetwork and the domain classification subnetwork, so that the gradient signal generated by the domain classification subnetwork during the optimization process is inverted by the gradient inversion layer and then transmitted to the shared feature extraction subnetwork, driving the shared feature extraction subnetwork to learn feature representations that eliminate frequency domain differences; The domain-invariant features are simultaneously input into the domain classification subnetwork and the task prediction subnetwork. By minimizing the classification loss of the task prediction subnetwork and maximizing the domain confusion loss of the domain classification subnetwork, a feature mapping relationship is established that makes the signals of different frequency domains tend to be distributed and aligned in the feature space, thereby realizing the adaptive transfer and fusion of voiceprint features.
4. The method according to claim 1, characterized in that, A feature library is constructed using the fused feature vectors. The similarity between the voiceprint signal to be identified and the features in the feature library is calculated to obtain the feature similarity, which includes: A multidimensional feature library is constructed using the fused feature vectors. The multidimensional feature library organizes the fused feature vectors through a hierarchical index structure, wherein the hierarchical index structure divides the fused feature vectors into multiple feature clusters according to the degree of feature similarity. Hash functions are calculated for each of the multiple feature clusters based on the hierarchical index structure. The voiceprint signal to be identified is converted into a retrieval feature vector, and candidate feature clusters are determined through the hierarchical index structure. The hash function corresponding to the candidate feature cluster is invoked to map the retrieved feature vector to the multidimensional feature library, and the feature similarity is obtained by calculating the vector distance in the hash space.
5. The method according to claim 1, characterized in that, Noise reduction parameters are generated based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the speaker signal are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector, including: Noise reduction parameters are generated based on the feature similarity. The noise reduction parameters include a reconstruction vector and a modulation vector. The reconstruction vector and the modulation vector are mapped to the frequency domain and the amplitude domain, respectively, to form a dual-domain noise reduction strategy. The speaker signal is processed based on the dual-domain noise reduction strategy. The frequency features of the speaker signal are reconstructed using a reconstruction vector mapped to the frequency domain. The optimal reconstruction coefficient is calculated iteratively, and the frequency components are adjusted according to the optimal reconstruction coefficient. The amplitude features of the reconstructed signal are dynamically adjusted using a modulation vector mapped to the amplitude domain.
6. The method according to claim 1, characterized in that, Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated, and the noise reduction and enhancement result is fed back to the cross-frequency domain transfer learning network. The feature transfer strategy is optimized through the domain adversarial training method, including: Based on the adjusted voiceprint signal, a noise reduction and enhancement result is generated. The noise reduction and enhancement result is quantitatively evaluated by calculating the signal-to-noise ratio and spectral distortion, and optimization feedback information is generated. The optimized feedback information is fed back to the cross-frequency domain transfer learning network, and the network parameters are updated through the domain adversarial training method. The network parameter update process adopts an adaptive learning rate strategy, and the adaptive learning rate is dynamically adjusted according to the noise reduction effect evaluation results to optimize the feature transfer strategy.
7. An intelligent noise reduction processing system, used to implement the method of any one of claims 1-6, characterized in that, include: The acquisition module is used to acquire voiceprint signals, convert the voiceprint signals into time-frequency representation results, divide the time-frequency representation results into the frequency domain, and obtain multiple frequency domain signals. The transformation module is used to transform the multiple frequency domain signals respectively, obtain multi-scale feature vectors, construct a cross-frequency domain transfer learning network, input the multi-scale feature vectors into the cross-frequency domain transfer learning network to obtain fused feature vectors, and the cross-frequency domain transfer learning network constructs the feature mapping relationship between frequency domains through domain adversarial training to realize adaptive transfer and fusion of voiceprint features. The calculation module is used to construct a feature library using the fused feature vector, calculate the similarity between the voiceprint signal to be identified and the features in the feature library, and obtain the feature similarity. An adjustment module is used to generate noise reduction parameters based on the feature similarity, the noise reduction parameters including a reconstruction vector and a modulation vector; the frequency features of the voiceprint signal to be identified are reconstructed and mapped using the reconstruction vector, and the amplitude features of the reconstructed signal are dynamically adjusted using the modulation vector; The feedback module is used to generate a noise reduction result based on the adjusted voiceprint signal to be identified, and to feed the noise reduction result back to the cross-frequency domain transfer learning network to optimize the feature transfer strategy through the domain adversarial training method.
8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.