A top-down multi-level prior information guided adaptive filtering method and system

Abstract

The application discloses a top-down multi-stage prior information guiding adaptive filtering method and system, and belongs to the technical field of speech signal processing. The method constructs a multi-stage neural processing framework, wherein each stage injects a prior information tensor of the same dimension through a unified addition mechanism, and the tensor is obtained through adaptive fusion calculation based on source-related prior characteristics and array observation shallow characteristics. The fusion strategy includes addition fusion, gate fusion and attention fusion, wherein the gate fusion generates a gate value through splicing and activation, and the attention fusion adopts an iterative refinement strategy to prevent local information loss. The network architecture takes SpatialNet as the backbone and combines Mamba structure to enhance time modeling. Through the consistent multi-stage injection mechanism, the application realizes explicit guidance and adaptive refinement of prior information, and effectively improves the robustness and performance of target speaker extraction.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing technology, specifically to a top-down, multi-level prior information-guided adaptive filtering method and system. Background Technology

[0002] In complex acoustic environments, extracting the speech signal of a target speaker using a microphone array is a crucial task in speech processing. Existing microphone array processing methods can be mainly divided into three categories: multi-channel filtering, source separation, and source extraction. Among them, source extraction aims to recover a specific target source signal.

[0003] Traditional source extraction methods are primarily based on a mask-based framework, isolating target source components by estimating and applying time-frequency or temporal masks. In these methods, prior information about the target speaker is encoded as an embedding vector, and the mixed speech representation is modulated through element-wise operations. However, the performance of such methods is fundamentally limited by the mask function form and the challenges of phase reconstruction. Subsequent temporal-domain models extend these concepts to learn masks in the latent representation space, but their expressive power is inherently limited by the quality of the original mixed speech representation because they are essentially performing a "subtractive" filtering on the observed mixed signal.

[0004] To overcome these limitations, research has shifted to a direct learning framework, aiming to learn a direct mapping from mixed speech to target speech. While this paradigm enables networks to reconstruct the target signal, challenges remain in effectively integrating target prior information. Existing methods typically emphasize the input dimension while neglecting the multi-level interaction between prior information and observations. Furthermore, traditional single-stage linear filtering methods perform filtering solely through the inner product of a linear filter and array observations, failing to fully utilize multi-level processing capabilities. Existing injection strategies lack a mechanism that maintains consistency across network stages and adaptively adjusts prior information based on observation features.

[0005] Therefore, the existing technology has the following main drawbacks: Existing methods often lack multi-level interaction when integrating prior information, resulting in insufficient utilization of prior information; The lack of a unified structural consistency mechanism makes it difficult to reuse and effectively use the same prior information across multiple network stages; There is a lack of effective fusion strategies that adaptively refine prior source information based on shallow features observed by arrays. Summary of the Invention

[0006] The purpose of this invention is to address the problems in existing technologies, such as the lack of multi-level interaction, poor structural consistency, and inability to adaptively refine prior information, by providing a top-down, multi-level prior information-guided adaptive filtering method and system to solve the aforementioned technical problems.

[0007] The technical solution of the present invention: In order to achieve the above-mentioned objective, according to a first aspect of the present invention, a top-down multi-level prior information-guided adaptive filtering method is proposed, comprising the following steps: Step 1: Construct a multi-stage neural processing framework: Based on the SpatialNet architecture, construct a multi-stage neural network containing N consecutive residual blocks. The input and output of each residual block maintain the same dimension. The input is a real tensor of the microphone array observation signal. Perform a short-time Fourier transform on the observation signals received by the M microphone arrays, and stack the real and imaginary parts of the complex spectrum of each channel along the channel dimension to form a real-valued observation tensor. Generate an initial hidden representation through the feature extraction block. Each stage in the network maintains a uniform dimension of the hidden variable tensor. The network structure adopts a similar architecture to SpatialNet as its backbone, where full-band blocks correspond to odd-indexed residual layers and narrow-band blocks correspond to even-indexed residual layers; the narrow-band blocks internally integrate the Mamba network structure.

[0008] The SpatialNet architecture consists of crossband blocks corresponding to odd-indexed residual layers and narrowband blocks corresponding to even-indexed residual layers. The narrowband blocks internally integrate the Mamba network structure to leverage its powerful temporal modeling capabilities.

[0009] Step 2, Define the prior feature tensor: Using an embedding extraction network, obtain source-related prior feature vectors associated with the target speaker; after performing a linear transformation on the source-related prior feature vectors, expand them into a prior feature tensor with the same dimension as the initial hidden representation through a broadcast operation. ; Step 3, Adaptive fusion calculation of prior information: Based on the prior feature tensor and the initial hidden representation, the prior information tensor is calculated through the adaptive fusion module. The prior information tensor is reused in each subsequent stage; the adaptive fusion module takes the prior feature tensor and the initial hidden representation as input, and generates the prior information tensor through one of the additive fusion strategy, gated fusion strategy, and attention fusion strategy. ; Step 4, Multi-stage prior information injection: Inject the prior information tensor The hidden representation of stage n is explicitly injected into the residual block output of each stage through addition operations. Update via the following methods: ,in Indicates the first One residual block; By using additive injection, the iterative refinement process is guided when prior information is consistent and scalable.

[0010] This injection method is similar to a structural bias, which constrains the output of each stage to the same prior guiding direction.

[0011] Step 5, Target Signal Recovery: The final hidden representation output from stage N is nonlinearly mapped. The signal is processed in blocks to obtain an estimate of its complex spectrum; an inverse short-time Fourier transform is performed on the estimated complex spectrum to synthesize a time-domain waveform, thereby obtaining the target speaker extraction signal.

[0012] In one possible embodiment, the specific process of forming the real-valued observation tensor in step 1 includes: transforming the time-domain observation signal received by an array of M microphones using uniform short-time Fourier transform parameters; wherein the frame length can be selected from 25ms to 32ms, the frame shift can be selected from 10ms to 16ms, the window function is a Hanning window or a Hamming window, and the FFT points are denoted as F. The normalized real and imaginary parts of the complex spectrum obtained from each channel are alternately arranged according to the channel dimension to form a three-dimensional real-valued observation tensor. Where T is the number of time frames, and 2M represents the sum of the real and imaginary channel numbers of the M microphones. This construction method fully preserves the spatial and phase information of the multi-channel signal.

[0013] In one possible embodiment, the feature extraction module consists of a cascaded convolutional layer, a normalization layer, and a nonlinear activation layer; this feature extraction block maps 2M input channels to C feature channels while keeping the number of time frames T and the number of frequency points F constant, i.e. The output of this feature extraction block is the initial hidden representation. This serves as the starting point for subsequent multi-stage processing.

[0014] In one possible embodiment, in step 1, N structurally isomorphic residual blocks are constructed. The residual blocks are n=1,2,…,N, and stacked in order; the internal processing flow for the nth residual block is as follows: First, input the real-valued observation tensor. This is achieved through a depthwise separable convolution module composed of channel-wise convolution and pointwise convolution, where the kernel size of the channel-wise convolution is [size missing]. (e.g., 3×3), the expansion rate can be gradually increased according to different stages to expand the receptive field, and point-by-point convolution is used for information interaction between channels; the output is divided into two parts along the channel dimension, and a nonlinear transformation of element-by-element multiplication is performed through gated linear units; then it is added to the input through residual connection to keep the output dimension C×T×F.

[0015] Therefore, all residual blocks maintain the same input and output dimensions, ensuring the consistency of the dimensions of the injected prior information.

[0016] In one possible embodiment, in step 2, the source-related prior feature vector is the speaker representation vector extracted from the target speaker's registered speech by a speaker embedding extraction network. The speaker embedding extraction network is a d-vector extraction network based on a time-delay neural network, an x-vector extraction network based on a residual network, or an ECAPA-TDNN network based on channel attention.

[0017] In one possible embodiment, in step 2, the linear transformation is to transform the source-related prior feature vector through a fully connected layer. Mapped to ; This represents the source-related prior feature vector. The dimension; the broadcast operation is to... The prior feature tensor is obtained by copying it T times along the time dimension and F times along the frequency dimension. .

[0018] In one possible embodiment, in step 3, the fusion strategy is an additive fusion, that is, directly summing the tensors: .

[0019] In one possible embodiment, in step 3, gate-based fusion involves dividing the broadcast features into parallel branches, the first branch being... Activation, second branch based and The splicing generation gating mechanism; the specific calculation formula is as follows: ;in This represents element-wise multiplication. This indicates a splicing operation.

[0020] In one possible embodiment, in step 3, attention-based fusion is performed: the prior feature tensor is... Consider it a query, a shallow feature Treat them as keys and values. Calculate the output using frequency-level or time-level cross-attention. To avoid the loss of local time-frequency information, an iterative refinement strategy is adopted to gradually fuse value tensors along the frequency and time dimensions in different blocks. .

[0021] In one possible embodiment, step 5 specifically includes the following process: The final hidden representation output at stage N. Input nonlinear mapping block The complex spectrum of the signal is estimated. An inverse short-time Fourier transform is applied to the obtained complex spectrum, with transform parameters strictly consistent with the forward transform in step 1, including frame length, frame shift, window function, and number of FFT points. The reconstructed time-domain frames are synthesized into a continuous time-domain waveform using the overlap-addition method, ultimately outputting the extracted signal of the target speaker.

[0022] In one possible embodiment, all trainable parameters in the method are jointly trained end-to-end, with the training objective being to minimize the scale-invariant signal-to-noise ratio loss between the estimated target signal and the real target signal.

[0023] According to a second aspect of the present invention, a multi-stage target speaker extraction system based on adaptive fusion of prior information is proposed to implement the above-mentioned multi-stage target speaker extraction method based on adaptive fusion of prior information, comprising: The observation tensor generation module is used to perform short-time Fourier transform on the observation signals received by M microphone arrays and stack the real and imaginary parts of the complex spectrum of each channel along the channel dimension to form a real-valued observation tensor. The feature extraction module is used to extract features from the real-valued observation tensor and generate an initial hidden representation with dimensions C×T×F; A multi-stage neural processing network consists of N structurally isomorphic residual blocks stacked sequentially, each residual block having the same input and output dimensions; Odd-indexed residual blocks are full-band blocks, and even-indexed residual blocks are narrow-band blocks, with the narrow-band blocks integrating a Mamba network structure; the multi-stage neural processing network processes the initial hidden representation stage by stage. The prior feature tensor generation module is used to obtain the source-related prior feature vectors of the target speaker using the embedding extraction network, and then perform linear transformation and broadcasting operations on the vectors to generate a prior feature tensor with the same dimension as the initial hidden representation. ; The adaptive fusion module uses the prior feature tensor Using the initial hidden representation as input, a prior information tensor is computed using additive fusion, gated fusion, or attention fusion strategies. ; The injection module injects the prior information tensor An addition operation is injected into the output of each stage residual block, updating the hidden representation of the nth stage to... ,in This represents the nth residual block; The signal recovery module is used to perform nonlinear mapping on the final hidden representation of the Nth stage to obtain the complex spectrum estimate of the signal, and synthesize the target speaker extracted signal through inverse short-time Fourier transform.

[0024] Compared with the prior art, the present invention has the following advantages: Adaptive refinement: By fusing shallow features to adaptively refine prior source information, the robustness of feature-level information integration is enhanced, effectively solving the problem of insufficient utilization of prior information.

[0025] Structural consistency: A unified injection mechanism ensures consistency in dimensions and physical meaning at each stage, simplifies architecture design, preserves physical interpretability, and avoids the expressive limitations of traditional masking methods. Flexibility: Three different fusion strategies (addition, gating, and attention) are provided, among which the gating-based fusion strategy shows the best performance balance in experiments and can be flexibly selected according to specific needs. Attached Figure Description

[0026] Figure 1 is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of a multi-stage neural network filtering framework. Figure (a) illustrates the source extraction method framework, with hidden representation... With prior information The dimensions are the same; Figure (b) shows the transformation process of the source embedding a; Figures (c), (d), and (e) show the additive fusion, gated fusion, and attention-based fusion structures, respectively; Figure 3 The results of the network output signals at each stage in the time-frequency domain. Detailed Implementation

[0027] 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 some embodiments of the present invention, but not all embodiments.

[0028] like Figure 1 As shown, a top-down, multi-level prior information-guided adaptive filtering method includes the following steps: Step 1: Construct a multi-stage neural processing framework to perform short-time Fourier transform, real-valued tensor quantization, shallow feature extraction and initial hidden representation generation on multi-channel observation signals, and build a multi-stage network main body composed of N cascaded residual blocks.

[0029] Step 2: Define the prior information tensor, extract source-related prior feature vectors from the registered speech, and obtain the prior feature tensor with the same dimension as the hidden representation through linear transformation and broadcast expansion.

[0030] Step 3: As Figure 2 As shown, the adaptive fusion computation of prior information generates a prior information tensor h that can be reused at each stage based on prior feature tensors and shallow features through addition, gating or attention fusion strategies.

[0031] Step 4: Multi-stage prior information injection. h is explicitly injected into the output of each residual block through addition operations to achieve iterative refinement.

[0032] Step 5: Target signal recovery, the final hidden representation output from stage N is nonlinearly mapped. The signal is processed in blocks to obtain an estimate of its complex spectrum; an inverse short-time Fourier transform is performed on the estimated complex spectrum to synthesize a time-domain waveform, thereby obtaining the target speaker extraction signal.

[0033] The following embodiments are all described in the order of the steps described above.

[0034] Example 1 This embodiment describes in detail the specific implementation of the multi-stage target speaker extraction method based on gated fusion.

[0035] Step 1: Constructing a multi-stage neural processing framework Speech was acquired using an array of M microphones (M=4 in this example). The sampling rate was 16 kHz, and mixed speech was acquired. STFTs were applied to each channel signal with a window length of 32 ms (512 points), a frame shift of 16 ms (256 points), and a Hanning window function, yielding the complex spectrum of each channel. The real and imaginary parts of each time-frequency unit were alternately stacked along the channel dimension to form a real-valued observation tensor. Where T≈250 (corresponding to 4 seconds of voice), F=257, 2M=8.

[0036] Construct feature extraction blocks Its specific structure is as follows: two one-dimensional convolutional layers with kernel sizes of 1×7 and 1×1 respectively, and an output channel count of L=128. Each convolutional layer is followed by batch normalization and a ReLU activation function. Input tensor After processing by feature extraction blocks, shallow features are obtained. And let the initial hidden representation Therefore .

[0037] Construct a network body consisting of N cascaded consecutive residual blocks, where the input and output dimensions of each residual block are both T×F×128. The odd-indexed residual blocks ( , , ) is a full-band cross block, retaining the original configuration of the source SpatialNet, and using a 1×5 convolution in the frequency dimension to achieve cross-frequency interaction; even-indexed residual blocks ( , , The residual blocks are narrowband blocks that integrate a bidirectional Mamba network structure for efficient long-range modeling with linear complexity along the time dimension. All residual blocks contain depthwise separable convolutions, gated linear units, residual connections, and layer normalization to maintain dimensionality consistency.

[0038] Step 2: Define the prior feature tensor The ECAPA-TDNN embedding extraction network is used to extract a 192-dimensional speaker embedding vector 'a' from a 2-second recording of the target speaker's registered speech. The network consists of: a feature extraction block (two 1D convolutional layers and three SE-Res2Blocks with dilation coefficients of 1, 2, and 3), a feature aggregation block (global average pooling to obtain a 3072-dimensional vector), and a non-linear projection block (two fully connected layers, outputting a 192-dimensional vector after L2 normalization).

[0039] The linear transformation of a is achieved through a fully connected layer. ,in , This yields a 128-dimensional alignment vector. Then, a broadcast operation is used to copy the alignment vector T times along the time dimension and F times along the frequency dimension to generate a prior feature tensor. This makes the 128-dimensional vector in each time-frequency unit equal to .

[0040] Step 3: Adaptive fusion computation of prior information (gated fusion) Prior feature tensor shallow features (Right now By concatenating the components along the channel dimension, we obtain the concatenated tensor. Input the spliced ​​tensor into the linear layer (1×1 convolution), reducing its dimension back to 128, then applying the Sigmoid activation function and layer normalization to obtain the gated signal; simultaneously, After individual Tanh activation and layer normalization, the final prior information tensor Calculated using the following gating fusion formula: .in, This represents element-wise multiplication. This indicates a splicing operation. It is the shallow feature output of the feature extraction block (and) same), Linear layers are used to reduce the dimensionality of the stitched data. Representation layer normalization, This represents the normalized Sigmoid activation function.

[0041] This gating mechanism enables the network to adaptively enhance or suppress the influence of prior information on each time-frequency unit based on the current observation characteristics. The dimension remains T×F×128 and is unchanged and reused in all subsequent stages.

[0042] Step 4: Multi-stage prior information injection For each stage n=1,2,…,6, the hidden representation is updated as follows:

[0043] in, This represents the nth residual block. Prior information tensor. As a structural bias, the output of each residual block is explicitly injected through addition operations, forcing the iterative process of each stage to always be constrained to the same target speaker guidance direction, effectively alleviating the problem of prior information decay in deep networks.

[0044] Step 5: Target signal recovery The final hidden representation output from stage 6 Feed into nonlinear mapping block The mapping block consists of two 1×1 two-dimensional convolutional layers: the first layer outputs 2M=8 channels, followed by PReLU activation; the second layer outputs 8 channels without activation, directly obtaining the estimated real and imaginary parts of the complex spectrum. An inverse short-time Fourier transform (iSTFT) is applied to the estimated complex spectrum, with transform parameters (window length, frame shift, window function, number of FFT points) strictly consistent with the forward STFT in step 1. Finally, the reconstructed time-domain frames are synthesized into a continuous time-domain waveform using an overlap-addition method, outputting the extracted signal of the target speaker.

[0045] Example 2 This embodiment adopts an attention-based fusion strategy and an iterative refinement strategy to address the problem of local time-frequency information loss.

[0046] The difference from Example 1 lies in the prior information tensor in step 3. Fusion computing.

[0047] In this embodiment, the prior features after broadcasting Treat it as a query, and treat shallow features It is considered as a key and a value.

[0048] An iterative cross-attention mechanism alternating between frequency-level and time-level attention is employed to compute the prior information tensor. To avoid losing local time-frequency information, an iterative refinement strategy is introduced to gradually fuse value tensors. The specific process is as follows: (1) For each and Perform linear projection and obtain query, key, and value tensors through 1×1 convolution: , ,

[0049] in, and For learnable 1×1 convolutions, the number of output channels remains at 128. The initial state of the values ​​is directly set to... To preserve the original observation information.

[0050] (2) Set the number of iterations R=2. For each round r=1,2, perform frequency-level cross-attention and time-level cross-attention sequentially: Frequency-level attention: For each time frame t, scaled dot product attention is computed in the frequency dimension while keeping t fixed.

[0051]

[0052] in, , .

[0053] Update frequency refinement value: .

[0054] All time frames are processed in parallel to obtain the intermediate value tensor. .

[0055] (3) After R iterations, the refined value tensor is obtained. The final prior information tensor By combining residual connections with initial shallow features, training can be stabilized and original information can be preserved.

[0056] The dimension remains T×F×128, and it remains fixed and reused in subsequent stages.

[0057] This attention fusion mechanism uses q as a constant query across time and frequency, which can selectively aggregate the components most relevant to the target speaker's identity from the observed features and gradually refine the fused representation through frequency-time alternation iteration.

[0058] Experimental verification The method of this invention was evaluated on the WSJ0-2mix dataset.

[0059] Hardware environment: Simulated four-sensor microphone array, with sensor positions randomly distributed within a sphere with a radius of 20cm.

[0060] Acoustic environment: The room impulse response was generated using the pyroomacoustics toolbox, and the reverberation time T60 was randomly sampled between 0.2s and 0.6s.

[0061] Data settings: Speech segments are divided into 4-second segments, and the registration sample duration is 2 seconds. The training set contains 20,000 samples, and the evaluation set contains 5,000 samples.

[0062] Compared to the baseline: The baseline method adds prior information directly to the shallow feature x_ only in the initial stage, instead of injecting it into each hidden variable as in this invention.

[0063] Evaluation metrics: Scale-invariant signal-to-noise ratio (SI-SDR), narrowband speech quality assessment (NB-PESQ), and number of model parameters.

[0064] Experimental Results Analysis: The time-spectrum diagram of the speech signal is as follows: Figure 3 As shown, the time-spectrum diagram of the speech signal gradually becomes clear at each stage, reflecting the consistency of the signal at each stage and the effectiveness of this prior information import.

[0065] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It is understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A top-down, multi-level prior information-guided adaptive filtering method, characterized in that, Includes the following steps: Step 1, Construct a multi-stage neural processing framework: Based on the SpatialNet architecture, construct a multi-stage neural network containing N consecutive residual blocks, where the input and output of each residual block maintain the same dimension; The input is a real tensor of the observation signals received by the microphone array. A short-time Fourier transform is performed on the observation signals received by the M microphone arrays. The real and imaginary parts of the complex spectrum of each channel are stacked along the channel dimension to form a real-valued observation tensor. The initial hidden representation is generated through feature extraction blocks, and the uniform dimension of the hidden variable tensor is maintained at each stage of the network. Step 2, Define the prior feature tensor: Using an embedding extraction network, obtain source-related prior feature vectors associated with the target speaker; after performing a linear transformation on the source-related prior feature vectors, expand them into a prior feature tensor with the same dimension as the initial hidden representation through a broadcast operation. ; Step 3, Adaptive fusion calculation of prior information: Based on the prior feature tensor and the initial hidden representation, the prior information tensor is calculated through the adaptive fusion module. The prior information tensor is reused in each subsequent stage; the adaptive fusion module takes the prior feature tensor and the initial hidden representation as input, and generates the prior information tensor through one of the additive fusion strategy, gated fusion strategy, and attention fusion strategy. ; Step 4, Multi-stage prior information injection: Inject the prior information tensor The hidden representation of stage n is explicitly injected into the residual block output of each stage through addition operations. Update via the following methods: ,in Indicates the first One residual block; Step 5, Target Signal Recovery: The final hidden representation output from stage N is nonlinearly mapped. The signal is processed in blocks to obtain an estimate of its complex spectrum; an inverse short-time Fourier transform is performed on the estimated complex spectrum to synthesize a time-domain waveform, thereby obtaining the target speaker extraction signal.

2. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 1, the specific process of forming the real-valued observation tensor includes: transforming the time-domain observation signal received by the array of M microphones using uniform short-time Fourier transform parameters; and alternating the normalized real and imaginary parts of the complex spectrum obtained from each channel according to the channel dimension to form a three-dimensional real-valued observation tensor. , where T is the number of time frames, and 2M represents the sum of the real and imaginary channels of M microphones.

3. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, The feature extraction module consists of a cascaded convolutional layer, a normalization layer, and a nonlinear activation layer. This feature extraction block maps 2M input channels to C feature channels, while keeping the number of time frames T and the number of frequency points F constant. The output of this feature extraction block is the initial hidden representation. This serves as the starting point for subsequent multi-stage processing.

4. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 1, N structurally isomorphic residual blocks are constructed. The residual blocks are n=1,2,…,N, and stacked in order; the internal processing flow for the nth residual block is as follows: First, input the real-valued observation tensor. The system employs a depthwise separable convolution module composed of channel-wise convolution and pointwise convolution, wherein the kernel size of the channel-wise convolution is [missing information]. The expansion rate can be gradually increased according to different stages to expand the receptive field. Pointwise convolution is used for information interaction between channels. The output is divided into two parts along the channel dimension and transformed by element-wise multiplication through gated linear units. Then, it is added to the input through residual connection to keep the output dimension C×T×F.

5. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 2, the source-related prior feature vector is the speaker representation vector extracted from the target speaker's registered speech by the speaker embedding extraction network. The embedding extraction network is a d-vector extraction network based on a time-delay neural network, an x-vector extraction network based on a residual network, or an ECAPA-TDNN network based on channel attention.

6. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 2, the linear transformation is to transform the source-related prior feature vector through a fully connected layer. Mapped to ; This represents the source-related prior feature vector. The dimension; the broadcast operation is to... The prior feature tensor is obtained by copying it T times along the time dimension and F times along the frequency dimension. .

7. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 3, the fusion strategy is an addition-based fusion, that is, directly summing the tensors: .

8. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 3, gating-based fusion involves dividing the broadcast features into parallel branches, with the first branch being... Activation, second branch based and The splicing generation gating mechanism; the specific calculation formula is as follows: ;in This represents element-wise multiplication. This indicates a splicing operation.

9. The top-down, multi-level prior information-guided adaptive filtering method according to claim 1, characterized in that, In step 3, attention-based fusion is performed: the prior feature tensor is... Considered a query, shallow features Treat them as keys and values; use frequency-level or time-level cross-attention to compute the output. .

10. A multi-stage target speaker extraction system based on adaptive fusion of prior information, used to implement the multi-stage target speaker extraction method based on adaptive fusion of prior information as described in any one of claims 1-9, characterized in that, include: The observation tensor generation module is used to perform short-time Fourier transform on the observation signals received by M microphone arrays and stack the real and imaginary parts of the complex spectrum of each channel along the channel dimension to form a real-valued observation tensor. The feature extraction module is used to extract features from the real-valued observation tensor and generate an initial hidden representation with dimensions C×T×F; A multi-stage neural processing network consists of N structurally isomorphic residual blocks stacked sequentially, each residual block having the same input and output dimensions; Odd-indexed residual blocks are full-band blocks, and even-indexed residual blocks are narrow-band blocks, with the narrow-band blocks integrating a Mamba network structure; the multi-stage neural processing network processes the initial hidden representation stage by stage. The prior feature tensor generation module is used to obtain the source-related prior feature vectors of the target speaker using the embedding extraction network, and then perform linear transformation and broadcasting operations on the vectors to generate a prior feature tensor with the same dimension as the initial hidden representation. ; The adaptive fusion module uses the prior feature tensor Using the initial hidden representation as input, a prior information tensor is computed using additive fusion, gated fusion, or attention fusion strategies. ; The injection module injects the prior information tensor An addition operation is injected into the output of each stage residual block, updating the hidden representation of the nth stage to... ,in This represents the nth residual block; The signal recovery module is used to perform nonlinear mapping on the final hidden representation of the Nth stage to obtain the complex spectrum estimate of the signal, and synthesize the target speaker extracted signal through inverse short-time Fourier transform.

Images