A voiceprint analysis method and system based on atlas recognition and a medium

Abstract

This application discloses a method, system, and medium for voiceprint analysis based on graph recognition, belonging to the field of voiceprint recognition technology. The method includes: acquiring voiceprint signals; generating Mel spectrum, bispectral amplitude spectrum, bispectral phase spectrum, and wavelet scale spectrum in parallel from the voiceprint signals; performing multi-scale peak detection on each modal spectrum, defining graph nodes, and constructing node features; constructing time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each modality, and generating subgraphs for each corresponding modality based on edge and node features; performing cross-modal semantic space alignment on each subgraph, constructing inter-modal correlation edges, and fusing them to obtain a heterogeneous topology fused voiceprint graph; and outputting noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression based on the heterogeneous topology fused voiceprint graph. This method can effectively preserve the time-frequency topology structure of the voiceprint and accurately identify noise types.

Description

Technical Field

[0001] This application belongs to the field of voiceprint recognition technology, and specifically relates to a voiceprint analysis method, system and medium based on graph recognition. Background Technology

[0002] With the continued deepening of global urbanization, the urban acoustic environment is becoming increasingly complex, and noise pollution has become one of the key issues affecting residents' quality of life and intelligent urban management. Various types of noise sources, such as traffic noise (vehicle horns, engine and tire friction noise), social noise (commercial advertising, crowd gathering noise), construction noise (decoration, building construction), and industrial equipment operating noise, are widely distributed in various functional areas of cities. The acoustic signature signals of these noise sources are non-stationary signals containing rich information, reflecting not only the physical characteristics of the sound source (such as frequency, energy, and harmonic structure) but also its spatiotemporal distribution, intensity evolution, and dynamic characteristics of interaction with the surrounding environment. Effectively analyzing these acoustic signatures has crucial engineering value and practical significance for achieving accurate noise source tracing, real-time monitoring, intelligent control, and building quiet and livable smart cities.

[0003] Currently, voiceprint analysis technology is mainly divided into two categories. The first is traditional noise monitoring technology, including fixed monitoring stations, portable devices, and manual patrols. This has problems such as limited coverage, high deployment costs, and slow response times, making it difficult to meet the needs of grid-based monitoring. The second is modern intelligent voiceprint analysis technology, which is mostly based on Euclidean space features such as Mel-frequency cepstral coefficients and short-time Fourier transform spectra, combined with models such as convolutional neural networks and recurrent neural networks to achieve noise classification. However, this type of method forcibly embeds the non-Euclidean structure of the voiceprint into the grid space, easily leading to inherent defects such as harmonic structure destruction, time-frequency manifold distortion, and confusion of similar noise types. Furthermore, it has insufficient environmental interference suppression capabilities and is difficult to adapt to the complex sound field environment of cities.

[0004] Meanwhile, topological modeling methods have begun to be adopted in the field of pattern recognition. Fingerprint recognition, palmprint recognition, iris recognition, and vein recognition technologies have made progress in recognition performance and cross-scenario applications by converting pattern features into graph manifold structures and using graph neural networks to model spatial topological relationships. However, the application of this topological modeling approach in voiceprint analysis is still in the exploratory stage, and further research is needed on techniques for fusing the time-frequency non-stationary characteristics of voiceprint signals with graph topological structures. Furthermore, existing voiceprint analysis patents often employ engineering optimization strategies involving high-dimensional feature extraction followed by compression, or rely on peripherals such as millimeter-wave radar to achieve multimodal fusion. These methods fail to reduce computational complexity at the feature representation level, resulting in high deployment costs and power consumption, making real-time inference difficult to achieve on resource-constrained edge nodes. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method, system, and medium for voiceprint analysis based on graph recognition. By utilizing parallel generation of multimodal graphs, heterogeneous topology fusion, graph neural network enhanced coding, and multidimensional collaborative attention mechanism, it can overcome the shortcomings of traditional Euclidean space feature modeling, improve noise recognition and localization accuracy, reduce computational complexity and power consumption at the edge, and achieve accurate voiceprint analysis in complex environments.

[0006] The following is the technical content of this disclosure: A voiceprint analysis method based on spectrum recognition includes: Acquire voiceprint signals; The Mel spectrum, bispectral amplitude map, bispectral phase map, and wavelet scaling spectrum are generated in parallel from the acoustic signal. Multi-scale peak detection is performed on each modal spectrum, graph nodes are defined, and node features containing frequency, time, energy, and harmonic correlation are constructed; Within each mode, temporal adjacency edges, frequency adjacency edges, harmonic correlation edges, and mutual information correlation edges are constructed. Based on the features of the edges and the nodes, subgraphs for each corresponding mode are generated. Cross-modal semantic space alignment is performed on each subgraph, intermodal association edges are constructed, and the heterogeneous topological fused voiceprint map is obtained by fusion. Based on the heterogeneous topology fusion voiceprint map, topology embedding features are extracted using graph neural network enhanced coding and multidimensional collaborative attention mechanism. Based on the topology embedding features, noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

[0007] Furthermore, The construction of time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each mode includes: Connect adjacent time frame nodes in the same frequency channel to construct temporal adjacency edges, and assign edge weights according to the differences in node energy and frequency; Connect adjacent frequency channel nodes in the same time frame to construct frequency domain adjacency edges, and assign edge weights according to node energy and frequency interval; For node pairs that satisfy the harmonic integer ratio constraint, construct harmonic association edges and assign edge weights according to frequency deviation, energy correlation, and phase coupling. For node pairs whose mutual information exceeds an adaptive threshold, construct mutual information-related edges and assign edge weights according to the mutual information values.

[0008] Furthermore, Cross-modal semantic space alignment is performed on each subgraph, intermodal association edges are constructed, and the resulting heterogeneous topological fused speaker graph is obtained, including: Graph encoding is performed on each modal subgraph, and node features are projected onto a unified manifold space; Calculate the feature cosine similarity and spatiotemporal correlation of each node in the subgraph of different modalities. For node pairs that meet the similarity and correlation thresholds, construct intermodal association edges and assign weights. Construct cross-scale association edges for nodes experiencing sudden noise; By fusing the subgraphs of each modality, the intermodal correlation edges, and the cross-scale correlation edges, a heterogeneous topological fused voiceprint map is obtained.

[0009] Furthermore, The process involves extracting topological embedding features based on the heterogeneous topological fusion acoustic signature, utilizing graph neural network enhanced coding and a multidimensional collaborative attention mechanism. Based on these topological embedding features, the system outputs noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and acoustic signature analysis data after environmental interference suppression. This includes: After performing multi-scale graph convolutional coding, graph attention enhancement, and graph Transformer coding on the heterogeneous topological fusion voiceprint map, topological embedding features are obtained by topological differential entropy pooling. Effective voiceprint features in topological embedding features are enhanced through node field attention mechanism, channel field attention mechanism, time-frequency field attention mechanism, and environment field attention mechanism; The lightweight edge inference engine is used to process the topological embedding features enhanced by the attention mechanism to obtain lightweight topological embedding features. Based on the lightweight topological embedding features, the noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

[0010] Furthermore, The noise type identification result based on topological embedding feature output, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression include: The topology embedding features are input into the noise type identification subnetwork, the sound source spatial localization subnetwork, and the noise parameter analysis subnetwork, which are set in parallel. The noise type identification results, sound source spatial localization coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output respectively.

[0011] Furthermore, The process of using a lightweight edge inference engine to process attention-enhanced topological embedding features includes: The attention-enhanced topological embedding features are sequentially subjected to single-layer linear dynamic manifold aggregation, topological dynamic pruning, and dynamic mixed precision quantization to obtain lightweight topological embedding features.

[0012] Furthermore, it also includes: Obtain feedback information from real-world scenarios; The weights of the multidimensional collaborative attention mechanism and the fusion weights of the heterogeneous topology fusion voiceprint map are updated based on feedback information from the actual scene, noise type identification results, spatial positioning coordinates of the sound source, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression.

[0013] A voiceprint analysis system based on graph recognition includes: The data acquisition module is used to acquire voiceprint signals; The topology graph construction module is used to generate Mel spectrum, bispectral amplitude spectrum, bispectral phase spectrum, and wavelet scale spectrum in parallel from the acoustic signal; perform multi-scale peak detection on each modal spectrum, define graph nodes and construct node features including frequency, time, energy, and harmonic correlation; construct time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each mode; generate subgraphs for each corresponding mode based on the edges and node features; perform cross-modal semantic space alignment on each subgraph, construct inter-modal correlation edges, and fuse them to obtain a heterogeneous topology fused acoustic graph. The graph analysis module is used to extract topological embedding features based on the heterogeneous topological fusion voiceprint graph, using graph neural network enhanced coding and multidimensional collaborative attention mechanism, and output noise type identification results, sound source spatial location coordinates, sound pressure level parameters and voiceprint analysis data after environmental interference suppression based on the topological embedding features.

[0014] Furthermore, The map parsing module is configured as follows: After performing multi-scale graph convolutional coding, graph attention enhancement, and graph Transformer coding on the heterogeneous topological fusion voiceprint map, topological embedding features are obtained by topological differential entropy pooling. Effective voiceprint features in topological embedding features are enhanced through node field attention mechanism, channel field attention mechanism, time-frequency field attention mechanism, and environment field attention mechanism; The lightweight edge inference engine is used to process the topological embedding features enhanced by the attention mechanism to obtain lightweight topological embedding features. Based on the lightweight topological embedding features, the noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

[0015] A computer storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method described thereon.

[0016] Compared with the prior art, this application has the following advantages: This scheme, which collects voiceprint signals and constructs a multimodal graph, extracts core feature nodes, establishes intramodal topological associations, and fuses cross-modal data to form a heterogeneous topological graph, then relies on graph neural networks and collaborative attention to complete feature extraction and multi-task parsing, uses non-Euclidean topological modeling to restore the true structure of voiceprints, fuses multi-dimensional complementary information, strengthens effective features, and reduces computational power consumption. Because this scheme avoids the defects of forced embedding in Euclidean space from the source of feature modeling, it fully preserves the harmonics and time-frequency topological characteristics of voiceprints, and greatly improves anti-interference and feature discrimination capabilities. Therefore, it can solve the technical problems of limited coverage, high cost, and slow response of traditional noise monitoring, as well as the existing voiceprint analysis features distortion, easy confusion of similar noise, and insufficient anti-interference.

[0017] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of the method of this disclosure is shown.

[0020] Figure 2 A schematic diagram of the system architecture of this disclosure is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Figure 1 A schematic diagram of the method of this disclosure is shown, and the specific implementation details of this disclosure include: 1. Preparatory stage: Acquire the voiceprint signal and preprocess it.

[0023] 2. Implementation Phase: S1: Steps for constructing a multimodal voiceprint topology map This step is the core innovation of this invention, which restores the voiceprint signal from the Euclidean grid feature space to the non-Euclidean topological graph space and realizes the structural-level fusion of multimodal graphs.

[0024] S11: Multimodal speaker map generation. Four complementary modal maps are generated in parallel from the preprocessed speaker signal x[n] (sampling rate 48kHz, 16-bit quantization).

[0025] 1) Mel spectrum: using adaptive window length short-time Fourier transform:

[0026] Window function The window is a Hanming window, and its length is based on the rate of energy change. Dynamic adjustment: transient segment steady state Gradual transition section After mapping with a 64-channel Mel filter bank, the Mel frequency conversion formula is:

[0027] Generate 64×T Vermeer spectra .

[0028] 2) Bispectral amplitude graph: based on third-order cumulants:

[0029] Finite sample estimation:

[0030] The bispectral was obtained by two-dimensional FFT:

[0031] Extracting amplitude With a frequency resolution of 93.75 Hz, a 128×128-dimensional bispectral amplitude map was generated. .

[0032] 3) Bispectral phase map: Extracting bispectral phase and phase coupling index

[0033] Generate phase map .

[0034] 4) Wavelet scaling spectrum: Continuous wavelet transform is performed using the Morlet wavelet basis:

[0035] Morlet wavelet function:

[0036] Scale sequence 64 samples were taken at equal intervals on a logarithmic scale, corresponding to the center frequency. Generate a 64×T dimensional wavelet scaling spectrum. .

[0037] S12: Time-Frequency Energy Peak Point Detection and Graph Node Definition Multi-scale peak detection based on the Hessian matrix is ​​performed on each modality spectrum. A Gaussian pyramid is constructed, with varying scales. :

[0038] Calculate the Hessian matrix at each point.

[0039] Eigenvalue decomposition ,when The value was initially identified as a local maximum. After cross-scale fusion, non-maximum suppression, and significance filtering, a set of peak points was obtained. Define graph nodes Construct 8-dimensional node feature vectors Normalized frequency Normalized time Logarithmic energy amplitude Local spectral centroid Spectral flatness Peak sharpness Spectral flux Harmonic correlation .

[0040] S13: Modal inner edge construction.

[0041] 1) Temporal adjacency edge: connects adjacent time frame nodes in the same frequency channel:

[0042] Edge weight:

[0043] 2) Frequency domain adjacency edge: connects adjacent frequency channel nodes in the same time frame.

[0044] Edge weight:

[0045] 3) Harmonic correlation edge: The frequency ratio satisfies the harmonic relationship. Construct edges from pairs of nodes:

[0046] Edge weight:

[0047] 4) Mutual information related edges: The Kraskov–Stögbauer–Grassberger K-nearest neighbor estimator is used. Estimating mutual information:

[0048] For node pairs exceeding the adaptive threshold (top 10% quantile globally), construct edges:

[0049] Edge weight: .

[0050] S14: Generation of independent subgraphs for each mode. Perform the above steps for each of the four modes to generate Mel spectrum subgraphs. Bispectral amplitude subplot Bispectral phase subgraph Wavelet scale subgraph .

[0051] S15: Cross-modal semantic space alignment and heterogeneous topology fusion graph generation. Lightweight graph encoding is performed on each subgraph, and a single-layer GCN extracts subgraph-level features.

[0052] Output Node features are projected onto a 64-dimensional unified manifold space using a modality-specific MLP.

[0053] Calculate the cosine similarity of projected features:

[0054] right And spatiotemporal correlation Node pairs with a time difference < 100ms and a frequency ratio < 1.2 are used to construct intermodal edges. Edge weights:

[0055] For sudden noise sources Dynamically constructing cross-scale related edges .

[0056] The final output is a heterogeneous topology fusion graph. The total number of nodes is 500–800.

[0057] Results: By constructing a multimodal acoustic topology map and explicitly modeling harmonic integer ratio constraints and time-frequency energy manifold structure, the accuracy of noise type identification reaches over 97%, and the accuracy of similar noise differentiation reaches over 95%, which is about 8% to 15% higher than the existing mainstream technologies; the sound source localization error is controlled within 3 meters, which can provide technical support for noise source tracing.

[0058] S2: Graph Neural Network Enhanced Encoding Steps S21: Multi-scale graph convolutional coding. Frequency-scale branches are used to divide the data into low [20, 500] Hz, mid [500, 2000] Hz, and high [2000, 8000] Hz subbands, with adjacency matrices designed for each. Convolution formula:

[0059] Time-scale branching is based on spectral flux characteristics Dynamic routing transient / steady-state branches:

[0060] Fusion output Stacked in 2 layers, the output node feature dimension is 256.

[0061] S22: Enhanced graph attention layer. Introducing 4-head self-attention:

[0062] Attention coefficients introduce edge-type embeddings :

[0063] The output dimension is 256 after being spliced ​​from 4 heads.

[0064] S23: Graph Transformer layer. Global self-attention:

[0065] Topological mask matrix If the shortest path distance is greater than 3, then... It employs a linearized Transformer (Performer, FAVOR+ mechanism).

[0066] in, complexity , The output node feature dimension is 128.

[0067] S24: Topological differential entropy pooling. Performs 4-scale processing on node features. Spectral clustering partitioning yields subgraph regions Calculate the differential entropy of each sub-graph region:

[0068] Scale weight Keep the one with the largest entropy value Subgraphs, summed pooling output:

[0069] S3: Multidimensional Collaborative Attention Enhancement S31: Node field attention. Node contribution weight:

[0070] Dynamic weights:

[0071] S32: Channel field attention. The 128-dimensional topological embedding vector is divided into 8 subgroups (16 dimensions per group), and the importance score for each group is calculated using SE-style:

[0072] in For ReLU, Dynamic recalibration The group weights are dynamically adjusted based on the preliminary predicted noise type.

[0073] S33: Time-frequency field attention. Define energy trajectory density:

[0074] Dynamically Adjusting the Time-Frequency Analysis Window Length: High Density + High Evolution Rate Employs a 4–8ms short window; low density + steady state A long window of 32–64ms is used; a medium window of 16ms is used for the gradient area.

[0075] S34: Environmental attention. Lightweight, deep separable convolutional CNNs extract environmental disturbance features. :

[0076] Interference weight:

[0077] Adaptive suppression:

[0078] Results: Relying on a multi-dimensional collaborative attention mechanism, it can separate the target voiceprint from background clutter. The recognition accuracy remains above 90% in low signal-to-noise ratio scenarios, which is about 18% higher than existing technologies. It can adapt to complex urban sound environments such as rainy days, strong winds, and multiple background clutter.

[0079] S4: Lightweight edge inference engine steps.

[0080] S41: Single-layer linear dynamic manifold aggregation. Composite dynamic normalized adjacency matrix:

[0081] Symmetric normalization Single-layer linear projection:

[0082] No iteration, no attention, no non-linear activation, sparsity >95%, complexity .

[0083] S42: Dynamic pruning of topology. Edge importance score:

[0084] The pruning rate is 65% in steady-state scenarios and 50% in dynamic scenarios.

[0085] S43: Dynamic Mixed Precision Quantization. Core Channel (Channel Field Attention Score) The core channels (approximately 30%) use 8-bit quantization; non-core channels (approximately 70%) use 2-bit or ternary quantization. Ternary quantization training:

[0086] Backpropagation. Combined with Huffman coding, the model volume compression rate is ≥94%.

[0087] S44: Heterogeneous computing task partitioning of MCU+FPGA. The MCU performs audio preprocessing, spectrogram generation, graph construction, attention calculation, and classification output; the FPGA performs sparse matrix operations for single-layer linear manifold aggregation.

[0088] S5: Steps for accurate multi-task voiceprint analysis.

[0089] S51: Noise type identification branch. Hierarchical classifier: 8 coarse classes and 40 fine subclasses.

[0090]

[0091] S52: Sound source spatial localization branch. GCC-PHAT time delay estimation:

[0092] Weighted least squares localization:

[0093] Weight Based on phase synchronization confidence, the positioning error is ≤3 meters.

[0094] S53: Noise parameter analysis branch. Synchronously outputs instantaneous sound pressure level and equivalent continuous A-weighted sound level. Cumulative percentage sound level The duration of the event and the sound pressure level error are ≤0.5dB(A).

[0095] Results: Employing multi-task collaborative analysis, it can simultaneously perform functions such as noise type identification, sound source localization, and sound pressure level monitoring, outputting analysis results at different granularities to meet the needs of smart city noise monitoring and management.

[0096] S6: Online adaptive optimization. Integrates an online error feedback mechanism to dynamically fine-tune the attention weights of S3 and the feature fusion weights of S1.

[0097] Closed-loop feedback and continuous optimization steps: S61: Hierarchical classification result output. Adaptively generates three levels of output: basic analysis results (noise type, intensity level, duration), refined analysis results (source coordinates, feature fingerprint, confidence level), and abnormal voiceprint warning results (sudden high decibels, rare noise).

[0098] S62: Output adaptation and low-power operation. Adapts to API interfaces, databases, and smart city management platforms, employing event-triggered and on-demand computing strategies.

[0099] S63: Closed-loop feedback and iterative system optimization. Continuously collect verification data and process feedback to maintain the manifold prototype center. .

[0100] Perform logarithmic power transformation on topological embedding features By compressing the dynamic range of energy and eliminating amplitude drift, and combining hierarchical manifold normalization to align the feature distributions of different time periods, the algorithm enhances feature stability and recognition consistency across time periods and scenarios.

[0101] Based on the above method, this application also provides a system, the system architecture of which is as follows: Figure 2As shown, it mainly includes the following core components: a data acquisition module, a topology graph construction module, a graph neural network enhanced coding layer, a multi-dimensional collaborative attention layer, a lightweight topology inference layer, a multi-task parsing layer, and an online feedback optimization layer. The voiceprint topology graph construction layer is responsible for fusing multimodal voiceprint spectra and constructing a heterogeneous topology fusion graph; the graph neural network enhanced coding layer performs deep feature learning on the topology fusion graph to extract topology embedding feature vectors; the multi-dimensional collaborative attention layer adaptively enhances effective voiceprint features and suppresses environmental interference; the lightweight topology inference layer enables low-power real-time inference at the edge; the multi-task parsing layer synchronously outputs noise type, location coordinates, and parameter information; and the online feedback optimization layer achieves adaptive model optimization based on feedback data. Specifically, it includes: The data acquisition module is used to acquire voiceprint signals; The topology graph construction module is used to generate Mel spectrum, bispectral amplitude spectrum, bispectral phase spectrum, and wavelet scale spectrum in parallel from the acoustic signal; perform multi-scale peak detection on each modal spectrum, define graph nodes and construct node features including frequency, time, energy, and harmonic correlation; construct time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each mode; generate subgraphs for each corresponding mode based on the edges and node features; perform cross-modal semantic space alignment on each subgraph, construct inter-modal correlation edges, and fuse them to obtain a heterogeneous topology fused acoustic graph. The graph analysis module includes: The graph neural network coding layer is used to perform multi-scale graph convolutional coding, graph attention enhancement and graph Transformer coding on heterogeneous topological fused voiceprint maps, and then obtain topological embedding features through topological differential entropy pooling. A multidimensional collaborative attention layer is used to enhance the effective voiceprint features in the topological embedding features through node field attention mechanism, channel field attention mechanism, time-frequency field attention mechanism, and environment field attention mechanism; The lightweight topology inference layer is used to sequentially perform single-layer linear dynamic manifold aggregation, dynamic topology pruning, and dynamic mixed-precision quantization on the attention-enhanced topology embedding features to obtain lightweight topology embedding features.

[0102] The multi-task parsing layer is used to input topology into the noise type identification subnetwork, sound source spatial localization subnetwork, and noise parameter parsing subnetwork, which are set in parallel. The outputs noise type identification results, sound source spatial localization coordinates, sound pressure level parameters, and voiceprint parsing data after environmental interference suppression, respectively.

[0103] Based on the same inventive concept, this disclosure also provides a computer storage medium having executable instructions stored thereon, which are executed by a processor to cause the processor to implement the above-described method.

[0104] Although this application 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A voiceprint analysis method based on atlas recognition, characterized in that, include: Acquire voiceprint signals; The Mel spectrum, bispectral amplitude map, bispectral phase map, and wavelet scaling spectrum are generated in parallel from the acoustic signal. Multi-scale peak detection is performed on each modal spectrum, graph nodes are defined, and node features containing frequency, time, energy, and harmonic correlation are constructed; Within each mode, temporal adjacency edges, frequency adjacency edges, harmonic correlation edges, and mutual information correlation edges are constructed. Based on the features of the edges and the nodes, subgraphs for each corresponding mode are generated. Cross-modal semantic space alignment is performed on each subgraph, intermodal association edges are constructed, and the heterogeneous topological fused voiceprint map is obtained by fusion. Based on the heterogeneous topology fusion voiceprint map, topology embedding features are extracted using graph neural network enhanced coding and multidimensional collaborative attention mechanism. Based on the topology embedding features, noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

2. The method of claim 1, wherein, The construction of time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each mode includes: Connect adjacent time frame nodes in the same frequency channel to construct temporal adjacency edges, and assign edge weights according to the differences in node energy and frequency; Connect adjacent frequency channel nodes in the same time frame to construct frequency domain adjacency edges, and assign edge weights according to node energy and frequency interval; For node pairs that satisfy the harmonic integer ratio constraint, construct harmonic association edges and assign edge weights according to frequency deviation, energy correlation, and phase coupling. For node pairs whose mutual information exceeds an adaptive threshold, construct mutual information-related edges and assign edge weights according to the mutual information values.

3. The method of claim 1, wherein, Cross-modal semantic space alignment is performed on each subgraph, intermodal association edges are constructed, and the resulting heterogeneous topological fused speaker graph is obtained, including: Graph encoding is performed on each modal subgraph, and node features are projected onto a unified manifold space; Calculate the feature cosine similarity and spatiotemporal correlation of each node in the subgraph of different modalities. For node pairs that meet the similarity and correlation thresholds, construct intermodal association edges and assign weights. Construct cross-scale association edges for nodes experiencing sudden noise; By fusing the subgraphs of each modality, the intermodal correlation edges, and the cross-scale correlation edges, a heterogeneous topological fused voiceprint map is obtained.

4. The method of claim 1, wherein, The process involves extracting topological embedding features based on the heterogeneous topological fusion acoustic signature, utilizing graph neural network enhanced coding and a multidimensional collaborative attention mechanism. Based on these topological embedding features, the system outputs noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and acoustic signature analysis data after environmental interference suppression. This includes: After performing multi-scale graph convolutional coding, graph attention enhancement, and graph Transformer coding on the heterogeneous topological fusion voiceprint map, topological embedding features are obtained by topological differential entropy pooling. Effective voiceprint features in topological embedding features are enhanced through node field attention mechanism, channel field attention mechanism, time-frequency field attention mechanism, and environment field attention mechanism; The lightweight edge inference engine is used to process the topological embedding features enhanced by the attention mechanism to obtain lightweight topological embedding features. Based on the lightweight topological embedding features, the noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

5. The method of claim 4, wherein, The noise type identification result based on topological embedding feature output, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression include: The topology embedding features are input into the noise type identification subnetwork, the sound source spatial localization subnetwork, and the noise parameter analysis subnetwork, which are set in parallel. The noise type identification results, sound source spatial localization coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output respectively.

6. The method of claim 4, wherein, The process of using a lightweight edge inference engine to process the attention-enhanced topological embedding features includes: The attention-enhanced topological embedding features are sequentially subjected to single-layer linear dynamic manifold aggregation, dynamic topological pruning, and dynamic mixed precision quantization to obtain lightweight topological embedding features.

7. The method of claim 1, wherein, Also includes: Obtain feedback information from real-world scenarios; The weights of the multidimensional collaborative attention mechanism and the fusion weights of the heterogeneous topology fusion voiceprint map are updated based on feedback information from the actual scene, noise type identification results, spatial positioning coordinates of the sound source, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression.

8. A voiceprint analysis system based on atlas recognition, characterized in that, include: The data acquisition module is used to acquire voiceprint signals; The topology graph construction module is used to generate Mel spectrum, bispectral amplitude spectrum, bispectral phase spectrum, and wavelet scale spectrum in parallel from the acoustic signal; perform multi-scale peak detection on each modal spectrum, define graph nodes and construct node features including frequency, time, energy, and harmonic correlation; construct time-domain adjacency edges, frequency-domain adjacency edges, harmonic correlation edges, and mutual information correlation edges within each mode; generate subgraphs for each corresponding mode based on the edges and node features; perform cross-modal semantic space alignment on each subgraph, construct inter-modal correlation edges, and fuse them to obtain a heterogeneous topology fused acoustic graph. The graph analysis module is used to extract topological embedding features based on the heterogeneous topological fusion voiceprint graph, using graph neural network enhanced coding and multidimensional collaborative attention mechanism, and output noise type identification results, sound source spatial location coordinates, sound pressure level parameters and voiceprint analysis data after environmental interference suppression based on the topological embedding features.

9. The system of claim 8, wherein, The map parsing module is configured as follows: After performing multi-scale graph convolutional coding, graph attention enhancement, and graph Transformer coding on the heterogeneous topological fusion voiceprint map, topological embedding features are obtained by topological differential entropy pooling. Effective voiceprint features in topological embedding features are enhanced through node field attention mechanism, channel field attention mechanism, time-frequency field attention mechanism, and environment field attention mechanism; The lightweight edge inference engine is used to process the topological embedding features enhanced by the attention mechanism to obtain lightweight topological embedding features. Based on the lightweight topological embedding features, the noise type identification results, sound source spatial location coordinates, sound pressure level parameters, and voiceprint analysis data after environmental interference suppression are output.

10. A computer storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 7.

Images