An AI music infringement identification method and system based on a voiceprint and a soundprint double engine

Abstract

This invention provides an AI-based music copyright infringement identification method and system based on a dual-engine approach using voiceprint and audioprint, relating to the field of audio content recognition and copyright protection. Its features include: source separation of the input audio to be detected, obtaining a dry vocal track and an accompaniment track; extraction of the singer's identity features from the dry vocal track using a voiceprint engine to form a voiceprint feature vector; and multi-dimensional feature analysis of the accompaniment track using a voiceprint engine. The advantages of this invention are: effective identification of complex infringement forms such as key changes, variations, instrument cloning, vocal cloning, and accompaniment cloning, significantly reducing missed and false positives; generation of a quantitative evidence report containing multi-dimensional similarity and specific infringement locations, and blockchain-based evidence storage, providing judicial-grade, traceable judgment results; and utilization of federated learning technology to achieve efficient cross-platform comparison while protecting data privacy, balancing detection accuracy, judicial validity, and large-scale application needs.

Description

Technical Field

[0001] This invention relates to the field of audio content recognition and copyright protection technology, and in particular to an AI music infringement recognition method and system based on a dual engine of voiceprint and audioprint. Background Technology

[0002] With the widespread application of generative artificial intelligence technology in music creation, the problem of unauthorized use of others' voice data to train AI models has become increasingly prominent. Some AI systems use deep learning to capture large amounts of publicly available recordings of singers and musicians, generating "copycat" works that are highly similar to, or even indistinguishable from, their timbre and style. This not only infringes on the original author's voice rights, copyright, and personality rights, but also blurs the legal boundaries between originality and reproduction. This phenomenon has sparked widespread concern and heated debate within the music industry regarding intellectual property protection, the safeguarding of artists' rights, and the compliance of AI-generated content. The judicial system, including intellectual property courts, lacks professional third-party institutions with both voiceprint and audioprint dual-engine capabilities to provide identification and authentication support and services to relevant legal entities.

[0003] Current technologies primarily rely on traditional audio fingerprint recognition (such as YouTube Content ID), but they lack robustness against adversarial perturbations such as speed changes, pitch shifts, and noise additions. Furthermore, they lack a collaborative analysis mechanism for singer's voiceprint and accompaniment's voiceprint, making it difficult to comprehensively cover the dual risks of infringement of personality rights and copyright. In addition, existing methods cannot effectively decouple and analyze multi-dimensional musical elements such as melody, rhythm, harmony, instruments, and arrangement structure, leading to high rates of missed and false judgments in complex infringement scenarios involving AI fine-tuning and reconstruction. Moreover, the judgment results lack interpretability and legal validity. Summary of the Invention

[0004] The purpose of this invention is to provide an AI music infringement identification method and system based on a dual engine of voiceprint and audioprint, which solves the problems of insufficient robustness of existing audio recognition methods when facing AI-generated music, inability to collaboratively analyze the singer's voiceprint and accompaniment's audioprint, difficulty in decoupling multi-dimensional music elements, resulting in a high rate of missed or incorrect judgments of dual infringements of personality rights and copyrights, and a lack of interpretability and judicial validity in the judgment results.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] An AI-powered music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint analysis, characterized by the following steps:

[0007] The input audio to be detected is subjected to sound source separation to obtain the human voice track and the accompaniment track;

[0008] The singer's identity features are extracted from the raw human voice track using a voiceprint engine to form a voiceprint feature vector.

[0009] The accompaniment track is analyzed using a soundprint engine to perform multi-dimensional feature analysis. The multi-dimensional features include at least one or more of the following: melody outline, rhythm fingerprint, harmony graph, instrument soundprint, key change, and arrangement structure.

[0010] The voiceprint feature vector and the multi-dimensional features are input into a multimodal fusion model and weighted to obtain a comprehensive similarity score.

[0011] The infringement determination is completed by comparing the comprehensive similarity score with a preset threshold, and a quantitative infringement evidence chain containing specific infringing elements and similarity values ​​is generated.

[0012] Furthermore, it also includes a source determination step: using an audio generation detection model based on ResNet or HiFi-GAN to determine whether the audio to be detected is AI-synthesized music; if it is not AI-generated, the weight of the voiceprint features is reduced.

[0013] Furthermore, the voiceprint engine employs the ECAPA-TDNN model and combines it with Softmax and AAM-Softmax loss functions for optimization, in order to extract voiceprint dry voice feature vectors that represent the singer's identity.

[0014] Furthermore, the melody analysis in the soundprint engine uses the MAE-MelodyNet self-supervised pre-trained model to extract pitch sequences and models long sequence dependencies through the Transformer-XL architecture to support melody contour comparison in cross-tonality and variation scenarios.

[0015] Furthermore, the rhythm analysis in the soundprint engine uses a temporal convolutional network (TCN) to extract rhythmic features and combines it with the Attention-DTW algorithm to achieve flexible temporal alignment, in order to be compatible with infringement forms involving variable speed and dynamic time scaling.

[0016] Furthermore, the harmony analysis in the sound signature engine uses constant Q transform (CQT) to extract time-frequency information and constructs a harmony spectrum through a graph convolutional network (GCN) to quantify the similarity of harmonic directions between different works.

[0017] Furthermore, the soundprint engine uses the Wave-U-Net model to separate multiple instruments from the accompaniment track, and extracts the instrument soundprint for each separated instrument through comparative learning in order to accurately locate plagiarism behavior of specific instrument tracks.

[0018] Furthermore, it also includes a federated learning cross-database comparison step, which, while ensuring that the copyright data of each party does not leave the local database, collaborates with multiple distributed copyright nodes to complete the infringement search and verification, thereby improving the search coverage and protecting data privacy.

[0019] Furthermore, it also includes a blockchain evidence storage step, which writes the infringement determination result, multi-dimensional feature fingerprints, timestamps and quantitative evidence summaries into the blockchain to achieve the solidification of judicial-grade evidence that is tamper-proof, verifiable and traceable.

[0020] An AI music copyright infringement identification system based on a dual engine of voiceprint and audioprint: including: an audio track splitting module, used to separate the input audio into a human voice track and an accompaniment track;

[0021] The voiceprint recognition module is used to extract the dry voiceprint features of the singer from the human voice track;

[0022] The sound signature analysis module is used to analyze the accompaniment track in multiple dimensions, including melody, rhythm, harmony, instruments, and arrangement structure.

[0023] The multimodal fusion judgment module is used to fuse voiceprint dry voice features and multi-dimensional voiceprint features, calculate the comprehensive similarity, and make infringement judgments.

[0024] The federated learning retrieval module is used for distributed comparison across copyright databases while protecting privacy.

[0025] The evidence generation module is used to generate a quantitative evidence report containing fine-grained infringement location information;

[0026] The blockchain evidence storage module is used to solidify evidence of infringement on the blockchain.

[0027] The beneficial effects of this invention are as follows:

[0028] (i) Achieve full-process, high-precision automated infringement detection: The system can effectively cover the entire lifecycle of AI music from training data tracing, content generation to online dissemination, and automatically and accurately identify infringement, solving the problem of insufficient robustness of existing technologies when facing AI-generated music.

[0029] (II) Breaking through the detection bottleneck of complex infringement scenarios: By decoupling the analysis of human voice dry sound and accompaniment six dimensions (melody, rhythm, harmony, instrument, arrangement, and key change) through the dual engine of "voiceprint-soundprint", it can accurately identify complex AI variant infringement behaviors that are difficult to deal with by traditional methods, such as key change, variation, rhythm expansion and contraction, instrument cloning, and arrangement fine adjustment, and significantly reduce the rate of missed judgment and false judgment.

[0030] (iii) Provide interpretable and judicial-grade judgment results: The system not only outputs a comprehensive similarity score, but also generates a quantitative evidence report containing specific infringement locations (such as sections), similarity values ​​of each dimension, and feature maps. The evidence is stored through blockchain to ensure that the evidence chain is complete, tamper-proof, and traceable, which greatly enhances the interpretability and judicial validity of the judgment results.

[0031] (iv) Balancing efficiency, privacy, and large-scale application: By adopting a federated learning framework, distributed secure comparisons can be achieved across platforms and institutions without sharing original copyright data, thus protecting data privacy while improving retrieval coverage. Simultaneously, the system supports large-scale batch scanning, suitable for the daily content review needs of platforms such as short videos and streaming media, which handle tens of millions of items. Attached Figure Description

[0032] Figure 1 This is an architectural block diagram of an AI music copyright infringement identification system based on a dual engine of voiceprint and audioprint.

[0033] Figure 2 This is a flowchart illustrating the overall process of infringement identification and retrieval for this invention.

[0034] Figure 3 This is a block diagram of the six-dimensional analysis module of the soundprint engine of this invention.

[0035] Figure 4 This is the logic diagram for the dual-engine fusion determination of the present invention.

[0036] Figure 5 This is a flowchart of an AI music copyright infringement identification method based on a dual engine of voiceprint and audioprint. Detailed Implementation

[0037] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0038] like Figures 1 to 5 As shown, an AI-based music copyright infringement identification method based on a dual-engine approach using voiceprint and audioprint is characterized by the following steps: source separation is performed on the input audio to be detected to obtain a vocal track and an accompaniment track; the singer's identity features are extracted from the vocal track using an audioprint engine to form a voiceprint feature vector; multi-dimensional feature analysis is performed on the accompaniment track using an audioprint engine, wherein the multi-dimensional features include at least one or more of the following: melody outline, rhythm fingerprint, harmony spectrum, instrument voiceprint, key shift, and arrangement structure; the voiceprint feature vector and the multi-dimensional features are input into a multimodal fusion model for weighted calculation to obtain a comprehensive similarity score; the comprehensive similarity score is compared with a preset threshold to complete the infringement determination and generate a quantitative infringement evidence chain containing specific infringing elements and similarity values. Here, voiceprint refers to the individual vocal characteristics of the singer, used for personality rights identification.

[0039] Phonogram: refers to the set of features of all musical elements in the accompaniment audio except for human voice.

[0040] The system also includes a source identification step: using an audio generation detection model based on ResNet or HiFi-GAN to determine whether the audio to be detected is AI-synthesized music; if it is not AI-generated, the weight of the voiceprint features is reduced. The voiceprint engine uses the ECAPA-TDNN model and is optimized by combining Softmax and AAM-Softmax loss functions to extract the voiceprint dry voice feature vector representing the singer's identity. The melody analysis in the voiceprint engine uses the MAE-MelodyNet self-supervised pre-trained model to extract pitch sequences and models long sequence dependencies through the Transformer-XL architecture to support melody contour comparison in cross-tonality and variation scenarios. The rhythm analysis in the voiceprint engine uses a temporal convolutional network (TCN) to extract rhythmic features and combines the Attention-DTW algorithm to achieve temporal flexible alignment to accommodate infringement forms involving variable speed and dynamic time scaling.

[0041] In addition, the system includes a source determination module. This module uses a binary classification model based on ResNet or HiFi-GAN architecture to analyze the audio to be detected and output the probability that it is AI-generated music. The multimodal fusion determination module dynamically adjusts the weight of voiceprint features in the final comprehensive similarity calculation based on this probability. When the determination result is a high probability that it is not AI-generated (such as a human cover), the system will reduce the weight of voiceprint features to avoid misjudging infringement due to different singers, thus focusing more on the copyright comparison of the accompaniment.

[0042] The harmonic analysis in the soundprint engine uses constant Q-transform (CQT) to extract time-frequency information and constructs a harmonic spectrum through a graph convolutional network (GCN) to quantify the similarity of harmonic directions between different works. The soundprint engine employs a Wave-U-Net model to separate multiple instruments from the accompaniment tracks and extracts the instrumental soundprint for each separated instrument through comparative learning to accurately locate plagiarism on specific instrument tracks. It also includes a federated learning cross-database comparison step, which, while ensuring that copyright data from each party does not leave their local database, collaborates with multiple distributed copyright nodes to complete infringement retrieval and verification, thereby improving retrieval coverage and protecting data privacy. Furthermore, it includes a blockchain evidence preservation step, writing the infringement determination result, multi-dimensional feature fingerprints, timestamps, and quantified evidence summaries into the blockchain to achieve immutable, verifiable, and traceable judicial-grade evidence solidification.

[0043] An AI-powered music copyright infringement detection system based on a dual-engine approach of voiceprint and audioprint analysis comprises: an audio track splitting module for separating input audio into a vocal track and an accompaniment track; a voiceprint recognition module for extracting the singer's voiceprint features from the vocal track; a voiceprint analysis module for performing multi-dimensional feature analysis of the accompaniment track, including melody, rhythm, harmony, instruments, and arrangement structure; a multi-modal fusion judgment module for fusing voiceprint features with multi-dimensional voiceprint features, calculating a comprehensive similarity, and determining copyright infringement; a federated learning retrieval module for conducting distributed comparisons across copyright databases while protecting privacy; an evidence generation module for generating a quantitative evidence report containing fine-grained infringement location information; and a blockchain evidence storage module for storing and solidifying infringement evidence on the blockchain.

[0044] The federated learning retrieval module employs a horizontal federated learning architecture, where each copyright node shares global model parameters instead of raw audio features, and the FedAvg algorithm aggregates locally updated gradients in each iteration. A differential privacy noise reduction mechanism further enhances node data protection.

[0045] Audio Track Separation Module: This module utilizes algorithms such as Deep Clustering to accurately separate the input mixed audio into vocal and accompaniment tracks. Its core advantage lies in decoupling voiceprint and audiogram analysis, laying the foundation for subsequent dual-engine parallel processing. This effectively avoids interference between vocals and accompaniment, thus significantly improving the overall accuracy and robustness of infringement detection.

[0046] Voiceprint Recognition Module: This module employs an advanced ECAPA-TDNN model to extract the singer's unique identity features (such as spectral envelope and formant trajectories) from the separated vocal tracks, forming a highly discriminative voiceprint feature vector. Its advantage lies in its ability to accurately identify whether an AI-synthesized voice refers to a specific natural person, effectively addressing infringement scenarios such as "replacement of singers in cover versions," and enhancing robustness in noisy and complex environments through an optimized loss function.

[0047] The soundprint analysis module performs in-depth, multi-dimensional analysis of the separated accompaniment tracks, specifically including five aspects: melody outline, rhythm fingerprint, harmony spectrum, instrument soundprints, and arrangement structure. It utilizes a series of cutting-edge deep learning models such as MAE-MelodyNet, TCN, GCN, and Wave-U-Net to process different musical elements. Its core advantage is its ability to decouple the analysis of the deep structure of the accompaniment, accurately identifying complex AI-induced infringement behaviors such as key changes, variations, rhythmic stretching, instrument cloning, and arrangement plagiarism, overcoming the limitations of traditional methods that can only perform overall comparisons.

[0048] Multimodal fusion judgment module: As the core decision-making unit of the system, this module receives feature vectors from the voiceprint recognition module and a six-dimensional feature set from the voiceprint analysis module. It then performs weighted calculations using models such as the Multimodal Fusion Transformer to arrive at a comprehensive similarity score. Its advantage lies in its ability to collaboratively consider the dual risks of infringement of personality rights (voiceprint) and copyright (voiceprint), performing intelligent and quantitative infringement judgments (infringement / suspected / non-infringement) based on preset thresholds. This significantly reduces the error of subjective judgment and provides interpretable quantitative evidence for the final result.

[0049] Federated Learning Retrieval Module: This module is activated when an infringement is initially identified as "suspected." Utilizing a federated learning framework, it collaborates with multiple distributed copyright nodes to perform cross-database infringement retrieval and verification, ensuring that the original copyright data of all parties remains within their local databases. Its main advantage is that it significantly improves the coverage and accuracy of infringement retrieval while strictly protecting the data privacy and trade secrets of all participating parties, thus meeting data security compliance requirements.

[0050] Evidence Generation Module: This module is responsible for transforming the system's judgment results into legally valid evidence. It generates a structured, quantitative infringement evidence report, which not only includes the final infringement conclusion but also details the specific infringement location (e.g., measure), similarity values ​​for various dimensions (melody, rhythm, instruments, etc.), and related feature maps (e.g., melody outline comparison diagrams). Its core advantage is providing a fine-grained, interpretable, and objective judicial-grade chain of evidence, greatly enhancing the credibility of the judgment results and their effectiveness in judicial rights protection.

[0051] During the melody comparison process, the evidence generation module first extracts beat points and bar lines using Madmom or librosa, divides the melody sequence into bar units, and then performs local DTW alignment on the pitch sequence within each bar, thereby achieving bar-level infringement location.

[0052] During the melody comparison process, the beat points and bar lines are first extracted using Madmom or librosa to divide the melody sequence into bar units. Then, the pitch sequence within each bar is locally DTW aligned to achieve bar-level infringement location.

[0053] Blockchain Evidence Storage Module: This module writes key information such as infringement determination results, multi-dimensional feature fingerprints, timestamps, and quantified evidence summaries into the blockchain. Its fundamental advantage lies in leveraging the immutability, verifiability, and traceability of blockchain technology to solidify infringement evidence, ensuring the integrity and authenticity of the entire evidence chain, and providing solid and reliable technical support for potential subsequent legal proceedings or platform processing.

[0054] Example 1: AI-generated music key modulation fine-tuning infringement detection

[0055] The original song "X" is in C major, with the piano as the core accompaniment instrument and a fixed arrangement structure. Although the audio to be detected was transposed to G major, embellished, and slightly modified in terms of accompaniment structure and some instrument timbres to evade detection, the system still identified a high degree of similarity through multi-dimensional analysis. First, a deep clustering algorithm was used to separate the audio into vocals and accompaniment tracks; ECAPA-TDNN voiceprint analysis confirmed that the singer was different from the original singer (similarity only 0.35, far below the 0.85 threshold). Subsequently, MAE-MelodyNet combined with Transformer-XL was used for trans-tonal melody alignment, achieving a core melody skeleton similarity of 88.7%; TCN and Attention-DTW rhythm analysis showed a rhythmic pattern, tempo, and time value distribution similarity as high as 92.3%.

[0056] Harmonic spectrum comparison based on CQT and Graph Convolutional Network (GCN) showed that the harmonic progressions were completely identical (86.5% similarity); Wave-U-Net successfully extracted the piano track, with a voiceprint similarity as high as 91.5%, confirming that the AI ​​directly cloned the original piano performance; TSN network further revealed that the two works had a structural similarity of 85.2% in terms of section arrangement (intro, verse, chorus, etc.). Finally, the multimodal fusion judgment engine integrated data from various dimensions and concluded that the overall similarity far exceeded the preset high threshold. The system determined that the AI-generated music constituted infringement and automatically generated a quantitative evidence report containing various similarity indicators, melody alignment spectrum, and piano voiceprint comparison heatmap. Its key information (hash value, timestamp, summary) has been written into the blockchain, completing judicial-level evidence preservation.

[0057] Example 2: Combined copyright infringement detection of "replacement of cover singer + plagiarism of accompaniment"

[0058] When the system receives a cover song video uploaded by an amateur user, whose vocal style is significantly different from the original, but whose background accompaniment sounds highly similar to an unauthorized version of the popular song "Y," the system first performs a source determination step. Using a ResNet-based audio generation detection model, the system analyzes the audio and determines that it is sung by a real person (not AI-generated) with a 98% probability. Based on this result, the system dynamically adjusts its judgment strategy, significantly reducing the weight of voiceprint features in subsequent fusion calculations.

[0059] Subsequently, the system processed the mixed audio into separate tracks, separating the vocals from the accompaniment; voiceprint analysis confirmed that the singer was an unregistered amateur, ruling out the possibility of the original singer's involvement. Next, a deep comparison of the accompaniment was conducted using six-dimensional voiceprint analysis technology: the similarity of the main melody reached 82.1%, the similarity of the rhythm and groove was 89.4%, the harmony progression was completely consistent (90.2%), the similarity of the key synthesizer and bass timbres was 87.8% and 85.3% respectively, and the overall arrangement structure matched 88.0%.

[0060] Given that the overall score fell into the "suspected infringement" range, the system automatically triggered a federated learning verification mechanism. Without accessing the original data from other platforms, it collaborated with distributed copyright nodes to compare the fingerprint of the accompaniment track, confirming a high degree of consistency with the accompaniment track of the song "Y" registered on multiple platforms. Ultimately, the system determined that the video accompaniment track constituted an infringement of the copyright of "Y," immediately issued a warning to the platform's review system, implemented traffic restrictions or removal operations, and pushed a rights protection report containing the infringing video ID, similarity data across various dimensions, and a complete chain of evidence to the copyright holder to support their subsequent legal actions.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint, characterized in that, Includes the following steps: The input audio to be detected is subjected to sound source separation to obtain the human voice track and the accompaniment track; The singer's identity features are extracted from the raw human voice track using a voiceprint engine to form a voiceprint feature vector. The accompaniment track is analyzed using a soundprint engine to perform multi-dimensional feature analysis. The multi-dimensional features include at least one or more of the following: melody outline, rhythm fingerprint, harmony graph, instrument soundprint, key change, and arrangement structure. The voiceprint feature vector and the multi-dimensional features are input into a multimodal fusion model and weighted to obtain a comprehensive similarity score. The infringement determination is completed by comparing the comprehensive similarity score with a preset threshold, and a quantitative infringement evidence chain containing specific infringing elements and similarity values ​​is generated.

2. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 1, characterized in that, It also includes a source determination step: using an audio generation detection model based on ResNet or HiFi-GAN to determine whether the audio to be detected is AI-synthesized music; if it is not AI-generated, the weight of the voiceprint feature is reduced.

3. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 2, characterized in that, The voiceprint engine uses the ECAPA-TDNN model and is optimized by combining Softmax and AAM-Softmax loss functions to extract voiceprint dry voice feature vectors that represent the singer's identity.

4. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 3, characterized in that, The melody analysis in the soundprint engine uses the MAE-MelodyNet self-supervised pre-trained model to extract pitch sequences and models long sequence dependencies through the Transformer-XL architecture to support melody contour comparison in cross-tonality and variation scenarios.

5. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 4, characterized in that, The rhythm analysis in the soundprint engine uses a temporal convolutional network (TCN) to extract rhythmic features and combines it with the Attention-DTW algorithm to achieve flexible temporal alignment, in order to be compatible with infringement forms involving variable speed and dynamic time scaling.

6. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 5, characterized in that, The harmonic analysis in the sound signature engine uses constant Q transform (CQT) to extract time-frequency information and constructs a harmonic spectrum through graph convolutional network (GCN) to quantify the similarity of harmonic directions between different works.

7. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 6, characterized in that, The soundprint engine uses the Wave-U-Net model to separate multiple instruments from the accompaniment track, and extracts the soundprint of each separated instrument through comparative learning in order to accurately locate plagiarism behavior of specific instrument tracks.

8. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 7, characterized in that, It also includes a federated learning cross-database comparison step, which, while ensuring that the copyright data of each party does not leave the local database, collaborates with multiple distributed copyright nodes to complete the infringement search and verification, so as to improve the search coverage and protect data privacy; It also includes a blockchain evidence storage step, which writes the infringement determination result, multi-dimensional feature fingerprints, timestamps and quantitative evidence summaries into the blockchain to achieve the solidification of judicial-grade evidence that is tamper-proof, verifiable and traceable.

9. The AI ​​music copyright infringement identification method based on a dual-engine approach of voiceprint and audioprint as described in claim 8, characterized in that, include: The audio track splitting module is used to separate the input audio into a human voice track and an accompaniment track; The voiceprint recognition module is used to extract the dry voiceprint features of the singer from the human voice track; The sound signature analysis module is used to analyze the accompaniment track in multiple dimensions, including melody, rhythm, harmony, instruments, and arrangement structure. The multimodal fusion judgment module is used to fuse voiceprint dry voice features and multi-dimensional voiceprint features, calculate the comprehensive similarity, and make infringement judgments. The federated learning retrieval module is used for distributed comparison across copyright databases while protecting privacy. The evidence generation module is used to generate a quantitative evidence report containing fine-grained infringement location information; The blockchain evidence storage module is used to solidify evidence of infringement on the blockchain.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method as claimed in any one of claims 1-9.

Images