A song forgery detection method and device based on graph frequency domain perception and global-local multi-scale fusion enhancement

Abstract

A song voice forgery detection method and device based on graph frequency domain perception and global-local multi-scale fusion enhancement. It relates to the technical field of computer hearing and audio security, and aims at the problems of insufficient song structure relationship representation, weak multi-scale forgery trace perception ability, difficulty in balancing global dependence and local details, and insufficient detection robustness in complex scenes in existing song voice forgery detection. The following scheme is proposed: obtaining the song voice signal to be detected and preprocessing, constructing the radial basis function song graph topology, mapping the graph frequency domain to obtain the radial basis function song graph frequency domain feature, then performing multi-scale time-frequency feature extraction, effective area gate enhancement, global dependence enhancement, local frequency refinement, graph attention modeling and graph pooling aggregation to generate a true or false discrimination result. The present application is suitable for deep forgery song voice detection in song synthesis, song timbre conversion, virtual singer, music platform content review and audio security supervision.

Description

Technical Field

[0001] This involves the field of computer hearing and audio security technology, specifically the detection of spoofed singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement. Background Technology

[0002] With the development of AI-generated content technology, vocal synthesis technology, and voice conversion technology, it has gradually become possible to generate singing voices with specific singer timbre, singing style, and emotional expression using deep learning models. Existing vocal generation technologies can simulate real singing voices in terms of melody, rhythm, pitch, timbre, and vocal production, resulting in highly natural and similar vocals. These technologies have been applied to virtual singers, music creation, digital entertainment, and personalized voice conversion. Compared to early, mechanical, and obviously distorted synthesized audio, current deepfake singing voices exhibit higher realism in spectral continuity, pitch variation, rhythmic expression, and emotional expression. Ordinary listeners and even some automated detection systems find it difficult to accurately distinguish them from real singing voices.

[0003] In the field of audio forgery detection, existing research largely focuses on deep speech forgery detection. For example, in scenarios such as speaker-synthesized speech, speech-to-speech conversion, and replay attack speech, traditional audio representations are extracted, including time-domain waveforms, Mel-frequency cepstral coefficients, constant Q-transform features, fast Fourier transform features, and spectrogram features. These are then combined with convolutional neural networks, recurrent neural networks, Transformer structures, graph attention networks, or voiceprint anti-spoofing models to distinguish between genuine and fake audio. These methods have formed relatively mature technical routes in ordinary speech anti-spoofing tasks, capable of identifying forgery traces, spectral anomalies, phase distortion, or abnormal voiceprint consistency to a certain extent.

[0004] However, singing and ordinary speech differ significantly in acoustic structure. Ordinary speech typically focuses on semantic expression and speaker identity information, with a relatively limited range of pitch variation and a relatively stable rhythmic structure. Singing, on the other hand, encompasses multiple dimensions of information, including melody, rhythm, pitch, timbre, articulation, breath control, vibrato, transitions, and emotional expression. Its spectral structure is more complex, with more dramatic temporal variations. Furthermore, in actual songs, it may be affected by factors such as accompaniment mixing, reverb processing, equalization, compression, and post-production editing. Therefore, directly applying speech forgery detection methods to singing forgery detection scenarios can easily lead to problems such as insufficient adaptation to the structural characteristics of singing, inaccurate localization of forgery traces, and decreased generalization ability under cross-language or cross-synthesizer conditions.

[0005] Existing methods for detecting fake singing voices typically still rely on traditional audio feature extraction approaches, primarily focusing on the time-domain waveform or conventional spectral representation of the audio. For example, some methods characterize the short-time spectral features of singing voices using Mel-frequency cepstral coefficients, while others obtain frequency-domain information through Fast Fourier Transform or Constant Q Transform, before inputting these features into deep neural networks for classification. While these methods can reflect some spectral differences in singing voice signals, they are essentially still feature representations based on regular grids or local windows, making it difficult to fully describe the structural relationships between singing voice sampling points, time-frequency units, and local anomalous regions. For deeply hidden, discontinuously distributed anomalies in deep fake singing voices that appear only in local frequency bands or local time segments, traditional features often fail to form sufficiently sensitive discriminative representations.

[0006] Meanwhile, existing deepfake detection models often suffer from an imbalance in utilizing global and local information during the feature enhancement stage. On the one hand, vocal forgery traces may manifest in long-range dependencies, such as abnormal pitch transitions across syllables or phrases, abnormal timbre consistency, or unnatural rhythmic changes. On the other hand, forgery traces may also be concentrated in local frequency details, such as abnormal high-frequency components, discontinuous local harmonic structures, and subtle distortions in vocal texture. If the model only emphasizes local convolutional features, it may easily overlook long-range dependencies in the overall vocal structure; if the model only emphasizes global attention modeling, it may weaken its ability to perceive local frequency anomalies and fine-grained forgery traces. Therefore, how to simultaneously capture the global dependencies, local spectral details, and multi-scale forgery features of vocals within the same detection framework is key to improving the accuracy and robustness of vocal forgery detection.

[0007] Furthermore, in practical applications, vocal forgery detection needs to address complex scenarios involving different datasets, languages, singers, synthesizers, and mixed use of accompaniment and non-accompaniment. Existing detection methods often perform well when training conditions are close to testing conditions, but their performance tends to degrade in scenarios involving cross-language, cross-synthesizer, or accompaniment interference. This indicates that existing methods are insufficient in extracting the essential structural features of vocal signals, lack strong ability to fuse forgery traces at different scales, and their generalization ability in complex scenarios still needs improvement. Especially in accompanied vocals, the anomalous information of forged vocals may be masked by the accompaniment, reverb, or post-processing, further increasing the difficulty of detection.

[0008] In summary, existing technologies suffer from insufficient attention to the graph structure relationships and graph frequency domain features of vocal signals, making it difficult to fully characterize the structural relationships between vocal sampling points; traditional time-domain or frequency-domain features have limited ability to perceive subtle anomalies and multi-scale forgery traces in deepfake vocals; existing detection models struggle to simultaneously consider global long-range dependencies, local frequency details, and multi-scale time-frequency feature enhancements; and they lack sufficient detection accuracy and generalization robustness in complex scenarios such as cross-language, cross-synthesizer, and accompaniment interference. Summary of the Invention

[0009] To address the shortcomings of existing technologies, such as insufficient attention to the graph structure and frequency domain features of vocal signals, making it difficult to fully characterize the structural relationships between vocal sampling points; limited ability of traditional time-domain or frequency-domain features to perceive subtle anomalies and multi-scale forgery traces in deepfake vocals; difficulty in simultaneously considering global long-range dependencies, local frequency details, and multi-scale time-frequency feature enhancements; and insufficient detection accuracy and generalization robustness in complex scenarios such as cross-language, cross-synthesizer, and accompaniment interference, the technical solution provided by this invention is as follows: A method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement includes: The steps include acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data. The steps are as follows: constructing a vocal graph topology based on the vocal signal data, and determining the correlation weights between sampling points in the vocal graph topology based on radial basis functions to obtain a radial basis function vocal graph topology. The step of performing graph-frequency domain mapping on the topology of the radial basis function song graph to obtain the frequency domain features of the radial basis function song graph; The step of inputting the radial basis function vocal map frequency domain features into the front-end feature extraction network for feature encoding to obtain the initial vocal feature map; The steps include performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map to obtain intermediate enhanced features; The intermediate enhanced features are subjected to parallel global dependency enhancement and local frequency refinement, and the enhancement results are fused to obtain the fused enhanced features. The steps are to perform graph attention modeling and graph pooling aggregation on the fused enhanced features to obtain aggregated discriminative features; The steps for generating a true / false discrimination result of the detected singing signal based on the aggregated discriminative features.

[0010] Furthermore, in a preferred embodiment, the preprocessing includes format unification, sampling rate adjustment, invalid segment filtering, and amplitude normalization of the vocal signal to be detected, to obtain vocal signal data that meets the requirements for constructing the vocal graph topology.

[0011] Furthermore, in a preferred embodiment, the radial basis function vocal graph topology uses sampling points in the vocal signal data as graph vertices, determines the association weights based on the temporal distance between sampling points, and adjusts the degree to which the association weights decay as the temporal distance increases through a scaling control parameter.

[0012] Furthermore, in a preferred embodiment, the radial basis function song graph topology also uses a threshold parameter to determine the effectiveness of the association between sampling points. When the association weight between sampling points is lower than the threshold parameter, the connection between the corresponding sampling points is removed.

[0013] Furthermore, in a preferred embodiment, the frequency domain features of the radial basis function singing graph are obtained by performing feature decomposition on the graph adjacency relationship corresponding to the topological structure of the radial basis function singing graph to obtain a graph frequency domain mapping basis, and the singing signal data is generated by performing a graph Fourier transform based on the graph frequency domain mapping basis.

[0014] Furthermore, in a preferred embodiment, the intermediate enhancement features are obtained through multi-scale fusion and partially gated bottleneck convolution. Multi-scale fusion extracts time-frequency features under different receptive fields through multi-layer multi-path convolution, and partially gated bottleneck convolution enhances the features of the effective region by masking the effective region, compressing the channel through bottleneck convolution, and weighting the gate coefficients.

[0015] A singing voice forgery detection device based on graph frequency domain perception and global-local multi-scale fusion enhancement is also provided, including: A module for acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data; Based on the singing signal data, a singing graph topology is constructed, and the correlation weights between sampling points in the singing graph topology are determined based on the radial basis function, thus obtaining the module of the radial basis function singing graph topology. A module for obtaining the frequency domain features of the radial basis function song graph is obtained by performing graph frequency domain mapping on the topological structure of the radial basis function song graph; The module that inputs the frequency domain features of the radial basis function singing map into the front-end feature extraction network for feature encoding to obtain the initial singing feature map; A module for obtaining intermediate enhanced features is obtained by performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map; The intermediate enhanced features are subjected to global dependency enhancement and local frequency refinement in parallel, and the enhancement results are fused to obtain a module with fused enhanced features; A module is obtained by performing graph attention modeling and graph pooling aggregation on the fused and enhanced features to obtain aggregated discriminative features; A module that generates a true / false discrimination result for the detected singing signal based on the aggregated discriminative features.

[0016] A computer storage medium is also provided for storing a computer program, which, when read by the computer, executes the method.

[0017] A computer is also provided, including a processor and a storage medium, wherein the computer executes the method when the processor reads a computer program stored in the storage medium.

[0018] A computer program product is also provided, which, when executed, implements the method described.

[0019] Compared with the prior art, the advantages of the technical solution provided by the present invention are as follows: By standardizing the format, adjusting the sampling rate, filtering invalid segments, and normalizing the amplitude of the vocal signals to be detected, this approach allows vocal data from different sources, recording conditions, loudnesses, and lengths to enter a unified processing flow, reducing interference from non-spoofing factors on the detection results. Existing vocal spoofing detection methods often rely directly on the original audio or conventional spectral features, which are easily affected by factors such as volume differences, silent segments, and inconsistent sampling conditions. This solution, through preprocessing, ensures that subsequent graph topology construction and feature enhancement are based on relatively stable data, thus improving the consistency between model training and actual detection.

[0020] Representing the sampling points in a singing signal as graph vertices and constructing a singing graph topology based on the relationships between these sampling points allows singing detection to move beyond the traditional one-dimensional time-series signals or the regular grid representation of two-dimensional spectrograms. Instead, it incorporates the structural relationships between sampling points within the singing signal into the feature representation process. Compared to existing conventional features such as MFCC, FFT, and CQT, which primarily focus on local spectral or short-term energy changes, this feature can express the relationships between different time positions in the singing signal, making it easier for the model to identify problems such as abnormal timbre continuity, temporal transitions, and local structural breaks in fake singing.

[0021] Based on radial basis functions, the correlation weights between sampling points in the vocal graph topology are determined. This allows for the establishment of continuously decaying connections based on the temporal and positional distances between sampling points, maintaining strong correlations between closer sampling points and gradually weakening correlations between more distant ones. Since vocal performance features melodic variations, pitch leaps, and rhythmic fluctuations, simply establishing connections based on fixed neighborhoods or equidistant windows fails to accurately reflect its true structure. This scheme constructs a graph topology using radial basis functions, making the connections between vocal sampling points more consistent with actual acoustic variations, thereby improving the ability to distinguish between the natural continuity of genuine vocal performance and abnormal correlations in forged vocal performance.

[0022] By adjusting the scaling parameters to control the degree to which the association weights of sampling points decay with increasing temporal distance, the topology of the vocal graph can adapt to vocal signals with different rhythms, pitch changes, and singing styles. Existing fixed-window feature extraction methods often struggle to handle complex situations such as slow sustained notes, rapid transitions, vibrato, and cross-syllable connections. This scheme adjusts the association range through scaling parameters, ensuring that the graph topology covers necessary local temporal relationships while avoiding over-connection of distant, irrelevant sampling points, thereby improving the flexibility and robustness of the graph structure representation.

[0023] By using a threshold parameter to determine the validity of associations between sampling points and removing weak connections below the threshold, the interference of distant weakly correlated sampling points, noise fragments, or invalid information on the graph structure can be reduced, making the graph topology sparser and more concentrated in representing valid vocal relationships. Compared to including all sampling point relationships in the calculation, this design can reduce the computational burden and discrimination noise caused by invalid connections, making subsequent graph frequency domain mapping more prominent in terms of real and valid temporal associations and structural anomalies.

[0024] By performing graph frequency domain mapping on the topology of the radial basis function vocal graph, the vocal signal can be transformed from a traditional time-domain or ordinary frequency-domain representation into a graph frequency-domain representation that includes graph structural relationships. This allows the model to analyze the vocal signal in the graph frequency dimension. Existing audio forgery detection typically relies on traditional spectral transformations, whose frequency analysis is based on regular time or frequency axes, making it difficult to reflect the non-Euclidean structural relationships between sampling points. This scheme combines graph topology with frequency domain analysis through graph frequency domain mapping, revealing more hidden structural frequency anomalies in forged vocals.

[0025] By performing feature decomposition on the graph adjacency relationships corresponding to the topological structure of the radial basis function singing graph, and then performing a graph Fourier transform on the singing signal data based on the obtained graph frequency domain mapping basis, frequency domain features of the radial basis function singing graph can be generated. These features simultaneously contain spectral variation information of the singing signal and structural relationship information between sampling points. Compared with traditional time-domain waveform, Mel-frequency cepstral coefficient, constant Q-transform, and fast Fourier transform features, they are more suitable for expressing nonlinear temporal dependency anomalies and structured forgery traces in deeply forged singing, thereby improving the discriminative power of the detection features.

[0026] By inputting the radial basis function (RBF) vocal graph frequency domain features into the front-end feature extraction network for feature encoding, the backbone detection network can receive vocal representations containing both graph structure and frequency domain information from the initial stage, rather than just ordinary audio features. Existing voiceprint anti-spoofing models are mainly geared towards speech forgery detection tasks. When directly transferred to vocal scenarios, they easily overlook the structural differences brought about by the complex melody, rhythm, and accompaniment mixing in vocals. This solution improves the front-end input features, making the subsequent network more adaptable to vocal forgery detection scenarios.

[0027] By extracting time-frequency features across different receptive fields using multi-layer, multi-path convolution, this approach can simultaneously cover short-term local textures, medium-scale spectral variations, and large-scale vocal structure information. Anomalies in deepfake vocals may appear at different scales, such as local harmonics, high-frequency noise, pitch transitions, trans-syllable continuity, or phrase-level rhythmic changes. A single convolutional scale is insufficient to capture all these anomalies. This approach enhances the perception of forgery traces at different granularities through multi-scale fusion, reducing the risk of missed detections associated with fixed-scale feature extraction.

[0028] Multi-scale fusion employs convolutional kernels of varying sizes and progressively expands the receptive field through multiple convolutions, enabling a more continuous and richer receptive field coverage without significantly increasing model complexity. Compared to ordinary convolutional networks that passively expand the receptive field by stacking layers, this design actively extracts time-frequency features of the singing from multiple scales, allowing both subtle local anomalies and larger-scale structural anomalies to be extracted and fused, thereby improving the model's adaptability to complex forged singing.

[0029] By concatenating and fusing the outputs of multi-path convolutions, the time-frequency information of singing extracted at different scales can be integrated into a unified feature representation, avoiding the dispersion of multi-scale features or their retention only in different network layers. Although some existing multi-scale models can extract features at multiple scales, the lack of an effective fusion mechanism can easily lead to insufficient utilization of information at certain scales. This solution uses concatenation and fusion to enable spoofing-sensitive features from different receptive fields to participate in subsequent discrimination, thereby improving the completeness of feature representation.

[0030] By masking the effective regions in the initial vocal feature map or multi-scale vocal features and performing convolution calculations only on these effective regions, the model can more effectively focus on processing areas that may contain forgery traces, reducing the impact of silent segments, accompaniment interference, weak energy regions, or irrelevant frequency bands. Compared to the traditional approach of performing indiscriminate convolution on the entire feature map, this design strengthens the modeling of effective regions and improves the ability to locate forgery-sensitive areas, making it particularly suitable for vocal detection scenarios with mixed accompaniment or complex post-processing.

[0031] By compressing channels of effective region features through bottleneck convolution, redundant channels and computational load can be reduced while retaining the main discriminative information, making subsequent global dependency modeling and local frequency refinement more focused. Existing detection models tend to introduce a large amount of redundant information and increase computational burden when directly performing attention calculations or convolutional enhancements on high-dimensional features. This scheme compresses feature representation through a bottleneck structure, enabling the model to improve detection performance while maintaining computational efficiency.

[0032] By weighting and enhancing the feature map after bottleneck convolution using gating coefficients, this approach dynamically strengthens forgery-sensitive regions and suppresses weakly correlated regions based on the importance of the features themselves. Ordinary convolution struggles to distinguish the contribution of different regions to the distinction between genuine and fake features, easily mixing accompaniment components, normal singing textures, and forgery anomalies. This scheme assigns different response intensities to different regions through a gating mechanism, enabling more distinct discriminative representations of features such as local anomalies, spectral distortion, and structural breaks.

[0033] By concatenating multi-scale fusion with partially gated bottleneck convolution to form a multi-scale fusion and partially gated bottleneck convolution module, the time-frequency features of the singing voice can be extracted from multiple scales first, and then effective regions can be selected and enhanced from the multi-scale features. Compared with a single approach that only uses multi-scale convolution or only uses gated selection, this combination not only expands the search range of forgery traces, but also improves the response intensity of effective forgery regions, making the output intermediate enhanced features more suitable for subsequent global and local enhancement processing.

[0034] Parallel global dependency enhancement and local frequency refinement can simultaneously take into account both the overall structural relationships of the singing and local spectral details. Singing forgery may manifest as long-range dependency anomalies across syllables, phrases, and frequency bands, or as local harmonic discontinuities, high-frequency texture distortion, or low-frequency fundamental frequency anomalies. Existing single-path models often struggle to accommodate both types of information. This approach uses parallel enhancement to ensure that both global consistency and local detail anomalies can participate in the final judgment.

[0035] Global dual-path contrastive attention extracts global correlation information through two independent attention paths, enabling the capture of long-range dependencies in vocal features from different attention perspectives. Compared to single-path attention mechanisms, which are susceptible to redundant correlations, this design can model global relationships in different channels or feature subspaces, making it easier for the model to detect anomalies in overall timbre consistency, melodic continuity, and long-distance time-frequency correlations in forged vocals.

[0036] By using learnable differential coefficients to modulate the results of two independent attention paths, we can enhance the truly discriminative global attention response and suppress redundant correlations shared by genuine and fake singing. Ordinary self-attention often focuses on high-energy or strongly correlated regions, but these regions do not necessarily have discriminative value for fakery. This approach highlights the effective differences between the two paths through differential attention, allowing the global enhancement features to more effectively reflect the unique abnormal patterns of fake singing.

[0037] By extracting local correlations between adjacent time-frequency regions through local context modeling, the model's ability to perceive short-term details and neighborhood changes can be enhanced. Local problems in singing voice forgery often manifest as unnatural changes in certain syllable boundaries, vibrato segments, aerophone textures, or local frequency bands. These are easily averaged or ignored by relying solely on global modeling. This approach preserves the details of local changes through local context modeling, improving the ability to identify fine-grained anomalies.

[0038] By decomposing local features into low-frequency and high-frequency components and enhancing them separately, we can differentiate the processing of forgery traces in different frequency regions. The low-frequency region is usually related to the fundamental frequency, pitch, sound stability, and main acoustic structure, while the high-frequency region is more likely to reveal harmonic details, noise textures, and synthesis traces. Existing methods tend to weaken these differences when mixing frequency components. This scheme, through branch frequency perception enhancement, effectively expresses both low-frequency structural anomalies and high-frequency detail anomalies.

[0039] By splicing the enhanced low-frequency and high-frequency components, the discriminative information from different frequency branches can be reintegrated, allowing the local enhanced features to simultaneously contain stable low-frequency structural information and detailed high-frequency texture information. Compared to detection methods that only focus on a certain frequency range, this design avoids over-reliance on a single frequency band, enabling the model to obtain relatively complete local frequency discrimination criteria under different singers, vocal ranges, and accompaniment conditions.

[0040] By processing the stitched local frequency features through multi-scale dilated convolution, the local modeling range can be expanded while maintaining low computational cost, enabling the model to capture anomalous correlations between adjacent and spaced time-frequency regions. Ordinary local convolution has a limited receptive field, making it difficult to cover forgery traces spreading across adjacent frequency bands or short time segments. This approach enhances the perception of anomalous details across different ranges through multi-scale dilated convolution.

[0041] By utilizing convolution to restore the channel dimension of local enhanced features, the features processed by low-frequency branches, high-frequency branches, and multi-scale dilated convolutions can be kept consistent with the subsequent fusion structure, avoiding instability in information transmission due to channel variations. This process enables the local frequency refinement results to be smoothly fused with global enhanced features and original input features, thereby ensuring the integrity of feature transmission within modules and the stability of network training.

[0042] By fusing global, local, and intermediate enhancement features, long-range dependency information and local frequency detail information can be introduced while preserving the original effective information. Compared to directly replacing the original features with enhancement features, this fusion method can reduce the loss of effective basic acoustic information, enabling the model to utilize both overall vocal structural anomalies and local fake texture anomalies, thereby improving the stability and accuracy of detection results.

[0043] By preserving intermediate enhanced features before the parallel global-local vocal feature enhancement module through residual addition, the problems of information decay and training instability during deep feature enhancement can be alleviated. Existing complex enhancement modules may over-rewrite the original discriminative features while improving expressive power. This scheme preserves the original multi-scale effective region features through residual connections, making the enhancement process more stable and also helping the model maintain its generalization ability under different datasets and detection scenarios.

[0044] By fusing and enhancing the input feature graph attention layer to model high-order dependencies between nodes, we can further uncover the abnormal distribution relationships of forgery traces in the time-frequency graph structure. Anomalies in deepfake singing voices are often not single points, but are scattered across multiple time segments, multiple frequency bands, or multiple acoustic structures. The graph attention layer can adaptively aggregate relevant information based on node importance, enabling scattered anomalies to form more complete discriminative evidence.

[0045] By hierarchically aggregating and reducing the dimensionality of graph attention features through graph pooling, key nodes and structural information that contribute more to true / false classification can be preserved, while reducing the impact of redundant nodes or weakly correlated features on classification. Compared to directly inputting all graph node information into the classification layer, graph pooling can improve the compactness of feature representation, enabling the model to maintain good detection stability in scenarios with accompaniment interference, cross-language, and cross-synthesizer scenarios.

[0046] Based on aggregated discriminative features, this method generates a true / false judgment result for the detected singing signal. It can uniformly transform the aforementioned graph frequency domain features, multi-scale time-frequency features, effective region enhancement features, global dependency features, local frequency refinement features, and graph structure aggregation features into an output conclusion of whether the singing is genuine or deeply forged. Compared to detection processes that rely on manual feature design or staged discrimination, this solution achieves collaborative optimization of feature construction, feature enhancement, structural modeling, and classification through an end-to-end approach, improving detection efficiency and ease of practical application.

[0047] Training the detection architecture using both real and deeply forged vocal samples enables the model to learn the differences between real and forged vocals in graph structure, graph frequency domain, multi-scale time-frequency features, global dependencies, and local frequency details. Compared to methods that rely solely on traditional audio features or manually defined rules for detection, this training method allows the model to adapt to forged features under different generation algorithms, singers, and song conditions, thereby improving actual detection capabilities.

[0048] By optimizing network parameters using weighted cross-entropy loss, the training bias problem caused by class imbalance between real and fake singing samples can be alleviated. Singing fakery detection datasets often suffer from insufficient samples in one class or uneven sample distribution under different testing conditions. Ordinary cross-entropy can easily cause the model to favor the class with more samples. This approach adjusts the learning intensity through class weights, enabling the model to maintain high sensitivity to samples from a minority of classes.

[0049] This method is applicable to both a cappella pure vocals and complete songs with accompaniment, enhancing its application value in real-world music platforms, virtual singer content management, audio security auditing, and copyright protection. Many existing detection methods perform well with clean speech or a cappella audio, but their performance tends to degrade when faced with mixed accompaniment, post-processing, multilingual, and cross-synthesizer conditions. This solution improves detection accuracy and generalization robustness in complex application scenarios through graph frequency domain awareness and global-local multi-scale fusion enhancement.

[0050] It is suitable for deepfake voice detection in voice synthesis, voice timbre conversion, virtual singers, music platform content review, and audio security supervision. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall framework of a singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement; Figure 2 A schematic diagram of the MSFPG (Multi-Scale Fusion and Partial Gated Bottleneck Convolutional Module); Figure 3 This is a schematic diagram of the parallel global-local vocal feature enhancement module PGLSFE. Detailed Implementation

[0052] To make the advantages and benefits of the technical solution provided by the present invention clearer, the technical solution provided by the present invention will now be described in further detail with reference to the accompanying drawings, specifically: Implementation Method 1: This implementation method provides a singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement, including: The steps include acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data. The steps are as follows: constructing a vocal graph topology based on the vocal signal data, and determining the correlation weights between sampling points in the vocal graph topology based on radial basis functions to obtain a radial basis function vocal graph topology. The step of performing graph-frequency domain mapping on the topology of the radial basis function song graph to obtain the frequency domain features of the radial basis function song graph; The step of inputting the radial basis function vocal map frequency domain features into the front-end feature extraction network for feature encoding to obtain the initial vocal feature map; The steps include performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map to obtain intermediate enhanced features; The intermediate enhanced features are subjected to parallel global dependency enhancement and local frequency refinement, and the enhancement results are fused to obtain the fused enhanced features. The steps are to perform graph attention modeling and graph pooling aggregation on the fused enhanced features to obtain aggregated discriminative features; The steps for generating a true / false discrimination result of the detected singing signal based on the aggregated discriminative features.

[0053] The preprocessing includes format unification, sampling rate adjustment, invalid segment filtering, and amplitude normalization of the vocal signal to be detected, so as to obtain vocal signal data that meets the requirements for constructing the vocal graph topology.

[0054] The radial basis function vocal graph topology uses sampling points in the vocal signal data as graph vertices, determines the association weights based on the temporal distance between sampling points, and adjusts the degree to which the association weights decay as the temporal distance increases through a scale control parameter.

[0055] The radial basis function song graph topology also uses a threshold parameter to determine the effectiveness of the association between sampling points. When the association weight between sampling points is lower than the threshold parameter, the connection between the corresponding sampling points is removed.

[0056] The radial basis function vocal graph frequency domain features are obtained by performing feature decomposition on the graph adjacency relationship corresponding to the topological structure of the radial basis function vocal graph to obtain the graph frequency domain mapping basis, and then generating the vocal signal data by performing graph Fourier transform based on the graph frequency domain mapping basis.

[0057] The intermediate enhancement features are obtained through multi-scale fusion and partial gated bottleneck convolution. Multi-scale fusion extracts time-frequency features under different receptive fields through multi-layer multi-path convolution. Partial gated bottleneck convolution enhances the effective region features by masking the effective region, compressing the channel through bottleneck convolution, and weighting the gate coefficients.

[0058] A singing voice forgery detection device based on graph frequency domain perception and global-local multi-scale fusion enhancement is also provided, including: A module for acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data; Based on the singing signal data, a singing graph topology is constructed, and the correlation weights between sampling points in the singing graph topology are determined based on the radial basis function, thus obtaining the module of the radial basis function singing graph topology. A module for obtaining the frequency domain features of the radial basis function song graph is obtained by performing graph frequency domain mapping on the topological structure of the radial basis function song graph; The module that inputs the frequency domain features of the radial basis function singing map into the front-end feature extraction network for feature encoding to obtain the initial singing feature map; A module for obtaining intermediate enhanced features is obtained by performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map; The intermediate enhanced features are subjected to global dependency enhancement and local frequency refinement in parallel, and the enhancement results are fused to obtain a module with fused enhanced features; A module is obtained by performing graph attention modeling and graph pooling aggregation on the fused and enhanced features to obtain aggregated discriminative features; A module that generates a true / false discrimination result for the detected singing signal based on the aggregated discriminative features.

[0059] A computer storage medium is also provided for storing a computer program, which, when read by the computer, executes the method.

[0060] A computer is also provided, including a processor and a storage medium, wherein the computer executes the method when the processor reads a computer program stored in the storage medium.

[0061] A computer program product is also provided, which, when executed, implements the method described.

[0062] Implementation Method Two: This implementation method is a further detailed description of the technical solution provided in Implementation Method One, specifically: This embodiment provides a method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement, used to distinguish between real and fake singing voices. The method first converts the input singing voice signal into a graph topology that can characterize the relationships between sampling points, and establishes connection weights between sampling points based on radial basis functions. Then, it obtains the frequency domain features of the singing voice graph through graph Fourier transform. Subsequently, the frequency domain features of the singing voice graph are input into an end-to-end detection network, and the fake-sensitive features are enhanced at the network front end through a global-local multi-scale feature enhancement module. Finally, after graph attention modeling, graph pooling aggregation, and classification output, the method obtains the judgment result of whether the singing voice to be detected belongs to a real singing voice or a deeply fake singing voice.

[0063] The singing signal to be detected is used as input, and the singing signal to be detected is preprocessed to obtain singing signal data that meets the requirements of subsequent graph topology construction. The process involves reading the vocal signals to be detected, standardizing their format, adjusting the sampling rate, filtering invalid segments, and normalizing the amplitude. This ensures that vocal signals from different sources, with varying lengths and loudnesses can be converted into a unified data representation. For songs with accompaniment, the entire song can be used as input, or the main vocal segments can be truncated or segmented according to specific detection needs. The preprocessed vocal signals serve as the foundation for subsequent graph topology construction, enabling the subsequent feature extraction process to proceed under relatively consistent signal conditions.

[0064] Based on the preprocessed singing signal, a singing graph topology is constructed to obtain graph structure data that characterizes the temporal and spatial relationships between singing sampling points; The sampling points in the preprocessed vocal signal are used as vertices in a graph structure, and the relationships between different sampling points are used as edges. For any two sampling points, the strength of their relationship is determined by their temporal distance; the closer the temporal distance, the stronger the relationship; the farther the temporal distance, the weaker the relationship. In this way, the original one-dimensional vocal signal arranged in chronological order is transformed into a graph structure with node and edge connections, allowing the potential temporal dependencies, local continuity, and structural relationships in the vocal signal to be utilized by subsequent models.

[0065] The edge weights in the song graph topology are calculated based on radial basis functions, and weakly associated edges are removed by threshold constraints to obtain the song graph topology based on radial basis functions. Using radial basis functions to calculate the correlation strength between sampling points, a scale control parameter is used to adjust the rate at which the correlation between sampling points decays over time. This maintains higher connection weights for nearby sampling points and reduces the connection weights for distant sampling points as distance increases. Simultaneously, a threshold parameter is set to determine the validity of the correlation. When the correlation strength between two sampling points falls below this threshold, it is considered that there is no valid correlation between the two sampling points, and the corresponding connection is removed from the graph topology. Through these processes, the constructed vocal graph topology retains the effective temporal correlations within the vocal signal while reducing the interference of weak or invalid long-distance correlations on the detection results.

[0066] By performing graph frequency domain mapping on the vocal graph topology based on radial basis functions, the frequency domain features of the radial basis function vocal graph are obtained; Based on the topological structure of the vocal graph obtained using radial basis functions, corresponding graph adjacency relationships are formed, and eigenvalue decomposition is performed on the graph structure to obtain orthogonal eigenvector matrices representing the frequency basis of the graph structure. Subsequently, the inverse matrix of these orthogonal eigenvector matrices is used as the basis for the graph Fourier transform, converting the vocal graph signal from the traditional time-domain representation to the graph frequency-domain representation, generating radial basis function vocal graph frequency-domain features. These vocal graph frequency-domain features not only contain the spectral variation information of the vocal signal itself but also the structural relationship information between sampling points, providing a structured discriminative input for subsequent detection networks that differs from traditional waveform features, Mel-frequency cepstral features, and ordinary frequency-domain features.

[0067] The frequency domain features of the radial basis function vocal map are input into the front-end feature extraction network to obtain the initial vocal feature map; The generated radial basis function vocal graph frequency domain features are input into the front-end feature extraction module of the voiceprint anti-spoofing network. The front-end feature extraction module may include convolutional layers, residual blocks, or other neural network structures for audio feature extraction. Its function is to initially encode the vocal graph frequency domain features, obtaining an initial vocal feature map containing information on the vocal spectrum distribution, temporal variations, and graph frequency domain structure. This initial vocal feature map serves as input to the subsequent global-local multi-scale feature enhancement module, enabling the feature enhancement process to directly apply to the vocal representation that already contains graph frequency domain information.

[0068] The initial vocal feature map is input into a multi-scale fusion and partially gated bottleneck convolution module to obtain intermediate enhanced features after multi-scale time-frequency modeling and effective region focusing; The multi-scale fusion and partially gated bottleneck convolution module includes a multi-scale fusion submodule and a partially gated bottleneck convolution submodule. The multi-scale fusion submodule adopts a multi-layer convolutional structure and sets convolutional kernels of different sizes in different convolutional paths to cover receptive fields of different scales. The first convolutional layer extracts time-frequency features in different local ranges through convolutional kernels of multiple prime scales; the second convolutional layer further expands the receptive field coverage based on the multi-path output of the first layer, enabling features of different scales to be cross-combined; the third convolutional layer concatenates and fuses the aforementioned multi-path outputs to obtain multi-scale vocal features that simultaneously contain fine-grained local features, medium-scale texture features, and large-scale structural features.

[0069] After obtaining multi-scale vocal features, a partial gated bottleneck convolution submodule uses a mask to mark the effective regions in the feature map for computation, performing convolution only on the marked effective regions. Subsequently, bottleneck convolution compresses the channel dimension of the effective region features to reduce redundant information and computational load. Then, convolution generates gating coefficients, which are applied to the feature map after bottleneck convolution, resulting in stronger responses in forgery-sensitive regions and suppression of weakly correlated regions or invalid background. After processing by this module, the resulting intermediate enhanced features contain both time-frequency information at different scales and highlight the effective regions relevant to forgery detection.

[0070] The intermediate enhanced features are input into the parallel global-local singing feature enhancement module to obtain fused enhanced features that simultaneously contain global dependency information and local frequency detail information. The parallel global-local vocal feature enhancement module comprises a global dual-path contrastive attention submodule and a branch frequency-aware local refinement submodule, which process intermediate enhanced features in parallel. The global dual-path contrastive attention submodule maps intermediate enhanced features to two independent sets of query, key, and value representations, and calculates the global correlation between features in each of the two attention paths. Subsequently, a learnable differential adjustment coefficient is introduced to differentially process the global attention results obtained from the two attention paths, enhancing discriminative global dependencies and suppressing redundant or ineffective attention responses. After this processing, global enhanced features reflecting long-range vocal dependencies, timbre consistency, melodic continuity, and overall structural anomalies are obtained.

[0071] Meanwhile, the branch frequency-aware local refinement submodule first models the local context in the intermediate enhanced features using depthwise separable convolutions to extract local correlations between adjacent time-frequency regions. Subsequently, the local features are decomposed into low-frequency and high-frequency components according to their frequency characteristics, and these components are enhanced separately. The low-frequency components reflect anomalies in the fundamental frequency, pitch, and relatively stable acoustic structures, while the high-frequency components reflect anomalies in harmonics, texture, noise, and subtle synthesis traces. The enhanced low-frequency and high-frequency components are concatenated, and the local modeling range is further expanded using multi-scale dilated convolutions. Finally, convolutions are used to restore the channel dimensions, yielding the local enhanced features.

[0072] The global and local enhancement features are fused and then added to the intermediate enhancement features preceding the input to the parallel global-local vocal feature enhancement module, resulting in a fused enhancement feature. This process preserves the original effective features, global consistency information, and local anomaly-sensitive information, enabling subsequent detection networks to simultaneously utilize both overall vocal structural anomalies and local spectral detail anomalies for discrimination.

[0073] The fused and enhanced features are input into the graph attention layer to obtain graph attention features that contain high-order dependencies between nodes; The fused enhanced features, after global-local multi-scale enhancement, are input into the graph attention layer of the voiceprint anti-spoofing network. The graph attention layer adaptively aggregates the fused enhanced features based on the importance relationships between different nodes or time-frequency locations, enabling forgery traces at different locations to be associated through graph structure relationships. For anomalous information that may be scattered across different frequency bands, time segments, or musical phrases in deepfake singing, the graph attention layer can model high-order dependencies between nodes, allowing these anomalous information to form a more complete discriminative representation, thus obtaining graph attention features.

[0074] The graph attention features are input into the graph pooling layer for hierarchical aggregation and dimensionality reduction to obtain aggregated discriminative features; Graph pooling layers filter, compress, and aggregate node information in graph attention features, retaining key nodes and structural information that contribute significantly to distinguishing between real and fake singing voices, while reducing the impact of redundant nodes, invalid time-frequency regions, and weakly correlated features on the classification results. Through graph pooling, the originally high-dimensional and complex graph attention features are transformed into more compact aggregated discriminative features, which can centrally reflect the frequency domain structure of the singing voice graph, multi-scale forgery traces, global dependency anomalies, and local frequency anomalies.

[0075] The aggregated discriminative features are input into the fully connected layer and the output layer to obtain the forgery detection result of the singing voice to be detected; The aggregated discriminative features output from the graph pooling layer are input into the fully connected layer. The fully connected layer performs discriminative mapping on the aggregated discriminative features and generates binary classification probabilities through the output layer. The binary classification probabilities represent the likelihood that the detected singing voice belongs to a real singing voice or a deepfake singing voice. When the probability corresponding to a deepfake singing voice meets the preset discrimination condition, the detected singing voice is output as a fake singing voice; when the probability corresponding to a real singing voice meets the preset discrimination condition, the detected singing voice is output as a real singing voice. Through this end-to-end detection method, a complete detection process is realized, from singing voice signal input, graph frequency domain feature construction, global-local multi-scale enhancement, graph structure dependency modeling, to real / fake discrimination output.

[0076] During the training phase, real singing samples and deepfake singing samples are input into the above detection architecture, and the network parameters are optimized using weighted cross-entropy loss. Training samples include real singing samples and deepfake singing samples obtained through singing synthesis, vocal timbre conversion, or other generation methods. For pure singing scenarios, unaccompanied singing samples can be used for training; for scenarios with accompaniment interference, complete song samples with accompaniment can be used. During training, radial basis function singing graph frequency domain features, intermediate enhancement features, fusion enhancement features, graph attention features, and aggregate discriminative features are generated sequentially according to the aforementioned process. The network parameters are updated based on the difference between the prediction results obtained from the output layer and the real labels of the samples. When using weighted cross-entropy loss, weights are set according to the class distribution of real and fake singing samples, so that the class with fewer samples receives sufficient learning strength during training, thereby reducing the impact of class imbalance on the detection model.

[0077] In the application phase, the singing voice to be detected is input into the trained detection architecture, and the detection conclusion of whether it is a real singing voice or a deeply faked singing voice is output. The trained detection architecture can be deployed in music content review platforms, virtual singer content management systems, audio copyright protection systems, network audio security monitoring systems, or other application scenarios requiring the identification of fake singing voices. For user-uploaded or platform-collected singing voices, the system processes them according to the same preprocessing, graph topology construction, graph frequency domain feature generation, global-local multi-scale enhancement, and classification process, ultimately outputting the result indicating whether the singing voice is genuine or fake. This method can be used for both a cappella singing and complete songs with accompaniment, and can adapt to the needs of deep fake singing voice recognition across languages, singers, synthesizers, and under different post-processing conditions.

[0078] Implementation Method 3: This implementation method is described in detail with reference to the accompanying drawings. Specific embodiments are provided to further illustrate the technical solutions offered above. Specifically: Example 1: This example provides a method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement, such as... Figure 1 As shown, the method for detecting fake singing includes the following steps: Step 1: Design a frequency domain feature sensing module for vocal graphs based on radial basis functions; Step 2: Construct a global-local multi-scale feature enhancement module; Step 3: Establish an end-to-end forgery detection architecture for singing.

[0079] This embodiment leverages the advantages of graph signal processing technology and deep learning to construct an efficient deepfake singing detection framework, RBF-GLMFE. The framework first constructs a radial basis function-based graph topology (RBSGT) for singing, mapping the singing signal to the graph frequency domain to generate RBSGF features. These features are then input into the AASIST backbone network, with a global-local multi-scale feature fusion enhancement module (GLMFE) embedded at the front end. Finally, discrimination is completed through a graph attention layer, graph pooling layer, and classification layer. This embodiment effectively integrates graph frequency domain structural representation with multi-scale feature enhancement capabilities, significantly improving the ability to perceive subtle anomalies in deepfake singing, while also possessing strong adaptability to both a cappella and accompanied singing scenarios.

[0080] Example 2: This example further defines the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement described in Example 1. The method in step 1 for designing the singing voice graph topology based on radial basis functions and generating the singing voice graph frequency domain features through graph Fourier transform is as follows: Step 1.1: Represent the singing signal as a graph topology and construct a graph topology based on radial basis functions to characterize the temporal and spatial relationships of the singing signal; Step 1.2: Map the singing signal from the traditional time domain to the graph frequency domain through graph Fourier transform to obtain the radial basis function singing graph frequency domain features, thereby enabling more in-depth analysis and processing of the signal in the graph frequency dimension; In this embodiment, step 1.1 is further defined, specifically including: First convert the singing signal S For image signal Specifically: For a vocal signal with several sampling points, it is first preprocessed to become a discrete sampling sequence arranged in time order. Each sampling point in this discrete sampling sequence corresponds to the amplitude of the vocal signal at a certain sampling time. Therefore, a graph vertex can be created for each sampling point according to its arrangement on the time axis, and the amplitude of the sampling point can be used as the node feature of the corresponding graph vertex.

[0081] In other words, the original singing signal is a one-dimensional sampling sequence arranged continuously over time. When converting it into a graph signal, the value of each sampling point is not changed. Instead, each sampling point is transformed from "a position in the one-dimensional sequence" into "a node in the graph structure," and the amplitude of the sampling point is attached to the corresponding node as the signal information carried by that node. Thus, the singing signal is represented as a graph signal defined on multiple graph vertices.

[0082] After completing the above transformation, the association between graph vertices is determined based on the temporal distance between different sampling points. Sampling points with closer temporal positions usually have stronger associations, while sampling points with farther temporal positions have weaker associations. Subsequently, the edge weights between graph vertices are calculated based on radial basis functions, and weak association connections are removed using a threshold, thereby obtaining a singing graph topology that can characterize the strength of temporal associations between singing sampling points.

[0083] Therefore, the conversion process can be summarized as follows: first, the vocal signal is discretized into a sequence of sampling points arranged in time; then, each sampling point is mapped to a vertex in the graph; the amplitude of the sampling point is used as the vertex feature; and finally, the temporal positional relationship between the sampling points is used as the basis for subsequent edge connection construction and edge weight calculation. This converts the original one-dimensional vocal signal into a graph signal.

[0084] Reconstruct the graph topology.

[0085] Specifically: Design a graph topology structure based on RBF to construct a singing voice signal, used to simulate the strength of connections between different sampling points in the singing voice signal, thereby obtaining the corresponding graph adjacency matrix. Its elements can be represented as:

[0086] in, The adjacency matrix of the graph is represented. The Middle i line, number j The matrix elements corresponding to the columns. These are scale control parameters. ij express and The greater the temporal and spatial distance between sampling points, the weaker their correlation. This is a threshold parameter indicating whether two sampling points are correlated. When the temporal correlation between the two sampling points is lower than a certain threshold, the threshold is set to 0.5. If the correlation between two vertices is 1e-3, then it is considered that there is no correlation between them. After introducing the threshold parameter, the correlation between sampling points will gradually weaken as the time interval increases, which is closer to the characteristics of actual singing signals.

[0087] And graph adjacency matrix The relationship between them is that the graph adjacency matrix can be used to... This can be understood as a relational table used to record the connection relationships of all sampling points. The horizontal and vertical axes of this relational table are arranged according to the order of sampling points in the singing signal, with each horizontal position corresponding to one sampling point and each vertical position corresponding to one sampling point.

[0088] This is the table where the relationship is located at the [number]th [position].i line, number j The specific value of the column. This value is used to describe the column... i The sampling point and the first j Does the sampling point have a connection, and how strong is the connection?

[0089] When the i The sampling point and the first j When there is a valid correlation between the sampling points A value greater than zero indicates a connection between the two, and the larger the value, the stronger the connection. i The sampling point and the first j When the correlation between sampling points is lower than a set threshold A value of zero indicates that no connection is established between the two.

[0090] therefore, It is made by all The overall matrix formed by arranging the sampling points according to their numbers Used to represent the overall topological structure of the vocal graph; That is A specific element in the sample is used to represent the connection weight between a pair of sample points.

[0091] In this embodiment, step 1.2 is further defined, specifically including: Eigenvalue decomposition is performed on the adjacency matrix of a symmetric undirected graph to obtain an orthogonal eigenvector matrix. U .Will As a graph Fourier base in the frequency domain of the mapping graph .based on Image signal Performing a graph Fourier transform yields the graph spectrum in the frequency domain. That is, the radial basis function song graph frequency domain feature RBSGF.

[0092] Example 3: This example further defines the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement described in Example 1. The method for constructing the global-local multi-scale feature enhancement module in step 2 is as follows: Step 2.1: Introduce the multi-scale fusion and partially gated bottleneck convolution module MSFPG. Multi-scale fusion uses three convolutional layers, which cover receptive fields of different scales through multiple convolutional kernels of different sizes. The size of each convolutional kernel is determined by a specific combination of prime numbers. A partially gated bottleneck convolutional block is constructed. The part of the feature map participating in the calculation is selected by using a mask label. Through bottleneck convolution and gate mechanism, convolution is performed only on the effective region. Step 2.2: Construct the parallel global-local vocal feature enhancement module PGLSFE. PGLSFE consists of global dual-path contrastive attention and branch frequency-aware local refinement. The two work in parallel to capture global dependencies and local details at the same time, and obtain a more robust acoustic feature representation through a fusion mechanism.

[0093] like Figure 1 As shown, this invention designs enhancement schemes from two perspectives: multi-scale time-frequency features and global-local dependencies. MSFPG is used to provide multi-scale time-frequency modeling capabilities and dynamically adjust the receptive field in the effective region convolution operation; PGLSFE combines global dependency modeling and local frequency refinement to enhance the model's ability to perceive subtle anomalies in fake singing.

[0094] Example 4: This example further defines the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement described in Example 3. The method for constructing the multi-scale fusion and partially gated bottleneck convolutional module MSFPG in step 2.1 specifically includes: The Multi-Scale Fusion and Partially Gated Bottleneck Convolutional Module (MSFPG) comprises two parts: the Multi-Scale Fusion Submodule (MSF) and the Partially Gated Bottleneck Convolutional Submodule (PGBC). For example... Figure 2 As shown, the multi-scale fusion submodule MSF consists of three convolutional layers: The first convolutional kernel selects a set of prime numbers. in Indicates the first The kernel size of each path. This represents the set of kernel sizes used in the first layer of convolution.

[0095] This indicates the kernel size used in the first convolutional path of the first layer; This indicates the kernel size used in the second convolutional path of the first layer; Indicates the first layer K The kernel size used in a single convolution path; First floor The output of a convolution path is defined as

[0096] in For the first i Convolutional kernels for each convolutional path; It is the activation function ReLU; This is a batch normalization operation; Indicates the first i Output feature maps for each path.

[0097] The second convolutional kernel selects a set of prime numbers. This represents the set of kernel sizes for the second layer, which is the set of kernel sizes used by each convolution path in the second layer.

[0098] This indicates the kernel size used in the first convolutional path of the second layer; This indicates the kernel size used in the second convolutional path of the second layer; Indicates the second layer K The kernel size used in the two convolutional paths; To further expand the receptive field coverage, since the sum of two prime numbers can be even, combining two layers of convolutional kernels of prime size can generate possible even receptive fields without increasing model complexity. The output of the first layer is known to be... The output feature maps for each path are as follows: , , The convolution process of the second layer can then be described as follows:

[0099] in It is the first i path to the first j Convolution kernels for each path; It is the first j Output of the path.

[0100] The third convolutional kernel concatenates the outputs of the second multipath layer and processes them... Convolutional fusion, this layer uses a set of kernel sizes. , This represents the set of kernel scales for the third layer, which is the set of kernel scales used in the third layer to fuse the multipath outputs of the second layer. By adding smaller kernels, the final receptive field range is expanded from even numbers to all integers, resulting in the MSF output:

[0101] in It is the final output convolution kernel; This is the final feature map output. Through this step, all multipath features are weighted and fused together to generate the final output feature map.

[0102] The partially gated bottleneck convolutional submodule PGBC is used to focus on effective regions and suppress invalid backgrounds based on multi-scale features, thereby reducing invalid computation and improving the ability to locate fake features.

[0103] Assume the input feature map is X , mask M Used to mark the valid region. For each feature map location. Convolution is performed on the effective region with a mask of 1, and the resulting output feature map is: B The input is fed into a bottleneck convolution, and its channel count is further compressed.

[0104] After locating the effective region in the input feature map using a gating mechanism, a targeted convolution is performed on that region; the gating coefficients are then... Feature map after convolution with bottleneck Multiply to obtain the weighted feature map. .

[0105] By cascading MSF and PGBC as described above, the MSFPG module can achieve multi-scale time-frequency modeling and adaptively focus on forged sensitive features within the effective region, providing a more discriminative input representation for the subsequent parallel global-local enhancement module.

[0106] Example 5: This example further defines the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement described in Example 3. The method for constructing the parallel global-local singing voice feature enhancement module in step 2.2 specifically includes: like Figure 3 As shown, the parallel global-local vocal feature enhancement module consists of a global dual-path contrastive attention submodule (GDPCA) and a branch frequency-aware local refinement submodule (FALR) operating in parallel. The outputs of the two modules are fused and then added to the input features as residuals to obtain the module output. .

[0107] The Global Dual-Path Contrast Attention (GDPCA) submodule aims to capture long-range dependency patterns in vocal features. The output is obtained from MSFPG. Its corresponding time-frequency sequence is The Query-Key-Value is obtained through three sets of linear mappings. Q , K Evenly distributed along the channel dimension ,in For the query and key of the first path, This represents the Query and Key for the second path. Subsequently, scaled dot product attention is calculated separately. .

[0108] The specific calculation process is as follows: For the first attention path, the relevance between the query representation and the key representation of the first path is calculated. During relevance calculation, each query feature in the first path is sequentially taken as the current feature to be focused on, and a dot product operation is performed with each key feature in the first path to obtain the similarity score between the current query feature and each key feature. Subsequently, the similarity score is scaled according to the channel dimension of the key features to reduce the impact of excessively large dot product results on the attention distribution when the channel dimension is large. The scaled similarity score is then normalized to convert the relevance score corresponding to each position into attention weights, thus obtaining the scaled dot product attention for the first path. .

[0109] For the second attention path, the relevance between the query representation and the key representation of the second path is calculated. During relevance calculation, each query feature in the second path is sequentially taken as the current feature to be focused on, and a dot product operation is performed with each key feature in the second path to obtain the similarity score between the current query feature and each key feature. Subsequently, the similarity score is scaled according to the channel dimension of the key features, and the scaled similarity score is normalized to convert the relevance score corresponding to each position into attention weights, thereby obtaining the scaled dot product attention of the second path. .

[0110] To enhance contrast and suppress redundant attention, learnable difference coefficients are introduced. Its definition is:

[0111] in The learnable path adjustment vector; These are initialization constants used to stabilize training. The final differential attention weights are:

[0112] Finally, the global features are updated, and the global path is output. .

[0113] The Branch Frequency Aware Local Refinement (FALR) submodule is used to characterize the local context and explicitly enhance frequency-related information. First, the local context is modeled. Then, low-frequency and high-frequency components are decomposed separately, and the enhanced low-frequency and high-frequency components are concatenated. Further, a multi-scale dilated convolution operation is performed, and [the process is then described in the original text]. Convolution restores the channel dimension, yielding the local path output. The process involves feature fusion and residual concatenation, followed by channel-level concatenation. Finally, the concatenated channel features are added to the input feature residuals to obtain an output that includes both global consistency information and local anomaly sensitivity information.

[0114] Through the parallel enhancement and fusion of GDPCA and FALR, this module can simultaneously strengthen the long-range dependency modeling capability and local frequency detail representation capability of singing features, thereby improving the detection effect of singing forgery.

[0115] Example 6: This example further defines the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement described in Example 1. Step 3, establishing an end-to-end singing voice forgery detection architecture, specifically includes: Step 3.1: Use the AASIST voiceprint anti-spoofing network as the backbone network. The AASIST model consists of a front-end feature extraction module, multiple graph attention layers, graph pooling layers, fully connected layers, and an output layer. Input the RBSGF features generated in Example 2 into the AASIST front-end. Figure 1 As shown, the GLMFE module constructed in Example 3 is embedded at the output of each front-end residual block: first, the MSFPG module performs multi-scale time-frequency modeling and effective region focusing, and then the PGLSFE module performs global dependency modeling and local spectral refinement in parallel, outputting enhanced features to the next processing unit. This embedding strategy enables the network to strengthen forgery-sensitive features early in feature extraction. The enhanced features are modeled by the graph attention layer to model high-order dependencies between nodes, accurately capturing the abnormal distribution patterns of forgery traces in the time-frequency graph structure; the graph pooling layer performs hierarchical feature aggregation and dimensionality reduction; finally, the fully connected layer and the output layer generate binary classification probabilities to complete end-to-end discrimination.

[0116] Example 7: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in any one of Examples 1 to 6.

[0117] Example 8: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in any one of Examples 1 to 6.

[0118] Example 9: A computer program product, which, when executed, implements the method described in Example 1.

[0119] Example 10: This example provides an embodiment to explain Examples 1 to 9 above. The specific embodiment is as follows: This embodiment presents the following embodiments, which are used to explain and illustrate a singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement.

[0120] Example 1 specifically includes: This embodiment proposes a method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement, including the following steps: Step 1: Construct a frequency domain feature perception module for vocal graphs based on radial basis functions; The specific process of step 1 is as follows: Step 1.1: Represent the singing signal as a graph topology, constructing a graph topology based on radial basis functions to characterize the temporal and spatial relationships of the singing signal. For the singing signal, the vertices of the graph correspond to the sampling points of the singing, and the edge weights are calculated using radial basis functions to characterize the temporal and spatial relationships of the singing signal; Step 1.2: Map the vocal signal from the traditional time domain representation to the graph frequency domain using graph Fourier transform to obtain the radial basis function (RBF) frequency domain features of the vocal graph, thereby enabling deeper analysis and processing of the signal in the graph frequency dimension. Eigenvalue decomposition is performed on the vocal graph topology to obtain eigenvalue matrices and orthogonal eigenvector matrices. A graph Fourier basis is constructed based on the inverse of the orthogonal eigenvector matrix, and a graph Fourier transform is performed on the graphical signal to generate the vocal graph frequency domain feature RBSGF.

[0121] Step 2: Construct a global-local multi-scale feature enhancement module, which includes a multi-scale fusion and partially gated bottleneck convolution module MSFPG and a parallel global-local singing feature enhancement module PGLSFE; The specific process of step 2 is as follows: Step 2.1: Introduce the multi-scale fusion and partially gated bottleneck convolution module MSFPG. Multi-scale fusion uses three convolutional layers, which cover receptive fields of different scales through multiple convolutional kernels of different sizes. The size of each convolutional kernel is determined by a specific combination of prime numbers. A partially gated bottleneck convolutional block is constructed. The part of the feature map participating in the calculation is selected by using a mask label. Through bottleneck convolution and gate mechanism, convolution is performed only on the effective region. Step 2.2: Construct the parallel global-local vocal feature enhancement module PGLSFE. PGLSFE consists of global dual-path contrastive attention and branch frequency-aware local refinement. The two work in parallel to capture global dependencies and local details at the same time, and obtain a more robust acoustic feature representation through a fusion mechanism.

[0122] Step 3: Establish an end-to-end singing voice spoofing detection architecture. Input the frequency domain features of the singing voice image into the AASIST voiceprint anti-spoofing network, and enhance the features through the global-local multi-scale feature fusion enhancement module to complete the singing voice deep spoofing detection task.

[0123] The specific process of step 3 is as follows: Step 3.1: The vocal image frequency domain feature perception module is used as input, and the global-local multi-scale feature enhancement module is embedded into the AASIST model. The advantages of each module are utilized to enhance the model's ability to learn fake vocal features, thereby more effectively reflecting the difference between real and deeply fake vocals and completing the identification of real and deeply fake vocals. The AASIST voiceprint anti-spoofing network is used as the backbone network. AASIST includes a front-end feature extraction module, a graph attention layer, a graph pooling layer, a fully connected layer, and an output layer. After the front-end residual block output, it first passes through the MSFPG module for multi-scale time-frequency modeling, and then through the PGLSFE module for global-local feature enhancement. The graph attention layer enhances the dependencies between nodes, the graph pooling layer performs feature dimensionality reduction and aggregation, and finally, the fully connected layer and the output layer perform binary classification.

[0124] In this embodiment: The process of detecting spoofed singing is as follows: First, given the complex spectral structure and nonlinear temporal dependence of singing signals, this invention utilizes graph signal processing techniques to map the singing from the time domain to the graph frequency domain, revealing hidden structured anomalies through RBSGF features. Based on this, a designed GLMFE module is used to capture spoofing traces of different granularities using the multi-scale receptive field of MSFPG, and to process global long-range dependencies and local spectral details in parallel using PGLSFE. This strengthens the spoofing-related feature representation before deep feature extraction, ultimately achieving high-precision and robust deep spoofing detection of singing.

[0125] The effectiveness of the proposed singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement will be further illustrated below with reference to the accompanying drawings and experimental data.

[0126] 1. Experimental setup Step 1: Conduct experiments on the CtrSVDD and MSCD datasets. The datasets contain unaccompanied vocal types (V) and complete songs with added accompaniment (M). Training and testing were performed under both V and M conditions. MSCD includes complex scenarios such as multilingual Test1, Test2, and cross-synthesizer Test3.

[0127] Step 2: Data Preprocessing and Feature Extraction. Invalid audio is filtered out, and the volume is normalized. The RBSGF features described in Step 1 of this embodiment are extracted as network input.

[0128] Step 3: Model training parameters. The Adam optimizer is used, with the initial learning rate set to... The weight decay is set to The weighted cross-entropy loss (WCE) is chosen as the loss function to mitigate the impact of class imbalance. The hyperparameters mentioned above are the preferred settings. 2. Feature and Module Processing Flow The RBSGF features are input into the front-end feature extraction module. After residual block processing, the features are optimized and enhanced by the GLMFE module.

[0129] First, the features are fed into the MSFPG module. The multi-scale fusion submodule extracts features using convolutional kernels of different sizes. Subsequently, some gated bottleneck convolutional submodules utilize masks. and gating coefficient The features are weighted to obtain features that retain the effective region and have undergone channel compression. Next, features Enter the PGLSFE module. The global path calculates differential weights through dual-path contrastive attention. Update global features to obtain .

[0130] Local paths are decomposed using frequency-aware decomposition, which categorizes features into low-frequency components. and high frequency After enhancement and reconstruction, the results were obtained. Finally, by fusing global and local features and introducing residual connections, the enhanced output is obtained. This enhancement feature The graph attention layer of AASIST is then used for further processing.

[0131] 3. Performance Analysis The effectiveness of the method described in this embodiment is verified through the following performance analysis: (1) Feature selection experiment On the AASIST baseline model, the traditional time-domain waveform Singing Audio, Mel frequency cepstral coefficients (MFCC), constant Q transform (CQT), fast Fourier transform (FFT), vocal graph topology based on radial basis function (RBSGT), and the vocal graph frequency domain feature (RBSGF) based on radial basis function of this invention were compared.

[0132] As shown in Table 1, on the MSCD dataset, RBSGF achieved better EER performance in the development set Dev, as well as the multilingual test sets Test1, Test and the cross-synthesizer test set Test3. This indicates that graph frequency domain features can more effectively characterize the structured information of singing signals and improve robustness in cross-condition and cross-synthesizer scenarios.

[0133]

[0134] Furthermore, as shown in Table 2, on the CtrSVDD dataset, RBSGF also achieves the best results under both pure vocal (V) and mixed accompaniment (M) conditions, and its absolute difference in EER between the test set and the development set (Diff.) is smaller, indicating that the graph frequency domain features have a significant enhancing effect on generalization ability.

[0135]

[0136] (2) Module ablation experiment To verify the contribution of each submodule in GLMFE, ablation analysis was performed by removing MSFPG, removing PGLSFE, and removing both at the same time.

[0137] As shown in Table 3, on the MSCD dataset, the model's EER increased significantly after removing MSFPG; this performance degradation was more pronounced in the cross-synthesis Test3 scenario after removing PGLSFE; and the overall performance further decreased when both sub-modules were removed simultaneously.

[0138]

[0139] Furthermore, as shown in Table 4, on the CtrSVDD dataset, removing any submodule leads to a decrease in test set performance and an increase in the absolute difference of EER, Diff.; while the complete model achieves the lowest EER under both V and M conditions, and the absolute difference of EER, Diff. is smaller, indicating that the GLMFE design of this invention can enhance generalization ability while improving accuracy.

[0140]

[0141] (3) Comprehensive performance comparison experiment The RBF-GLMFE method of this invention is compared with existing mainstream methods, including two categories: voice spoofing detection methods and singing voice spoofing detection methods.

[0142] As shown in Table 5, on the MSCD dataset, the method of the present invention achieves the best or significantly better EER performance in the Dev, Test1, Test2 and cross-synthesizer Test3 scenarios; especially in the Test3 scenario, the method of the present invention can still maintain a relative advantage, indicating that it has a strong generalization ability for unknown generation algorithms.

[0143]

[0144] Furthermore, as shown in Table 6, on the CtrSVDD dataset, the method of the present invention outperforms the comparison method in both pure vocal (V) and mixed accompaniment (M) scenarios, and the absolute difference in EER between the test set and the development set is smaller, indicating that the method of the present invention not only has higher detection accuracy, but also has stronger generalization ability and resistance to accompaniment interference.

[0145]

[0146] In summary, this embodiment verifies the singing voice forgery detection method based on graph frequency domain perception and global-local multi-scale fusion enhancement proposed in this invention. By introducing graph topology construction and graph frequency domain feature representation based on radial basis functions, and combining it with the GLMFE module to achieve global-local multi-scale fusion enhancement, the accuracy and robustness of singing voice forgery detection are effectively improved, especially in cross-language, cross-synthesizer and accompaniment interference scenarios, it has stronger generalization ability.

[0147] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement, characterized in that, include: The steps include acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data. The steps are as follows: constructing a vocal graph topology based on the vocal signal data, and determining the correlation weights between sampling points in the vocal graph topology based on radial basis functions to obtain a radial basis function vocal graph topology. The step of performing graph-frequency domain mapping on the topology of the radial basis function song graph to obtain the frequency domain features of the radial basis function song graph; The step of inputting the radial basis function vocal map frequency domain features into the front-end feature extraction network for feature encoding to obtain the initial vocal feature map; The steps include performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map to obtain intermediate enhanced features; The intermediate enhanced features are subjected to parallel global dependency enhancement and local frequency refinement, and the enhancement results are fused to obtain the fused enhanced features. The steps are to perform graph attention modeling and graph pooling aggregation on the fused enhanced features to obtain aggregated discriminative features; The steps for generating a true / false discrimination result of the detected singing signal based on the aggregated discriminative features.

2. The method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement according to claim 1, characterized in that, The preprocessing includes format unification, sampling rate adjustment, invalid segment filtering, and amplitude normalization of the vocal signal to be detected, so as to obtain vocal signal data that meets the requirements for constructing the vocal graph topology.

3. The method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement according to claim 1, characterized in that, The radial basis function vocal graph topology uses sampling points in the vocal signal data as graph vertices, determines the association weights based on the temporal distance between sampling points, and adjusts the degree to which the association weights decay as the temporal distance increases through a scale control parameter.

4. The method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement according to claim 1, characterized in that, The radial basis function song graph topology also uses a threshold parameter to determine the effectiveness of the association between sampling points. When the association weight between sampling points is lower than the threshold parameter, the connection between the corresponding sampling points is removed.

5. The method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement according to claim 1, characterized in that, The radial basis function vocal graph frequency domain features are obtained by performing feature decomposition on the graph adjacency relationship corresponding to the topological structure of the radial basis function vocal graph to obtain the graph frequency domain mapping basis, and then generating the vocal signal data by performing graph Fourier transform based on the graph frequency domain mapping basis.

6. The method for detecting fake singing voices based on graph frequency domain perception and global-local multi-scale fusion enhancement according to claim 1, characterized in that, The intermediate enhancement features are obtained through multi-scale fusion and partial gated bottleneck convolution. Multi-scale fusion extracts time-frequency features under different receptive fields through multi-layer multi-path convolution. Partial gated bottleneck convolution enhances the effective region features by masking the effective region, compressing the channel through bottleneck convolution, and weighting the gate coefficients.

7. A singing voice forgery detection device based on graph frequency domain perception and global-local multi-scale fusion enhancement, characterized in that, include: A module for acquiring the vocal signal to be detected and preprocessing the vocal signal to obtain vocal signal data; Based on the singing signal data, a singing graph topology is constructed, and the correlation weights between sampling points in the singing graph topology are determined based on the radial basis function, thus obtaining the module of the radial basis function singing graph topology. A module for obtaining the frequency domain features of the radial basis function song graph is obtained by performing graph frequency domain mapping on the topological structure of the radial basis function song graph; The module that inputs the frequency domain features of the radial basis function singing map into the front-end feature extraction network for feature encoding to obtain the initial singing feature map; A module for obtaining intermediate enhanced features is obtained by performing multi-scale time-frequency feature extraction and effective region gating enhancement on the initial singing feature map; The intermediate enhanced features are subjected to global dependency enhancement and local frequency refinement in parallel, and the enhancement results are fused to obtain a module with fused enhanced features; A module is obtained by performing graph attention modeling and graph pooling aggregation on the fused and enhanced features to obtain aggregated discriminative features; A module that generates a true / false discrimination result for the detected singing signal based on the aggregated discriminative features.

8. A computer storage medium for storing computer programs, characterized in that, When the computer program is read by the computer, the computer executes the method of claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1.

10. A computer program product, as a computer program, is characterized by: When the computer program is executed, it implements the method of claim 1.

Images