A method and system for cross-channel spoofed speech denoising and tracing

By constructing a residual orthogonal computation-based voiceprint tracing module and an autoencoder noise decoupling training framework, the problem of cross-channel forged voice tracing was solved, enabling accurate tracing of the real speaker in forged voices and improving the accuracy and reliability of voice evidence collection.

CN120126502BActive Publication Date: 2026-06-12ZHEJIANG UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
ZHEJIANG UNIV
Filing Date
2025-04-02
Publication Date
2026-06-12

AI Technical Summary

Technical Problem

Existing technologies cannot effectively trace the source of spoofed voice, especially in cross-channel transmission environments, where the voiceprint features of spoofed voice signals are sparse and highly coupled with the voiceprint of the target speaker, making it difficult to trace the source.

Method used

A voiceprint tracing module based on residual orthogonal computation is constructed. Combined with the noise decoupling training framework of autoencoder, the original voiceprint information of suspected voice forgers is extracted through confusion feature elimination and channel noise suppression. A generalized inference system is constructed to achieve cross-channel tracing.

Benefits of technology

It has achieved accurate extraction of the voiceprint features of the real speaker in fake voice recordings, improving the accuracy of voice evidence collection and the reliability of practical applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120126502B_ABST
    Figure CN120126502B_ABST
Patent Text Reader

Abstract

The application discloses a cross-channel fake voice noise reduction tracing method and system, and relates to the technical field of artificial intelligence. A voiceprint tracing module based on confusion feature elimination is constructed, confusion features of a target speaker voiceprint are eliminated through residual orthogonal calculation, and original voiceprint information of a fake voice suspect is extracted; a joint training framework oriented to channel noise decoupling is built, the noise separation capability of an autoencoder and the robustness of a voiceprint tracing model are utilized to suppress the interference of cross-channel noise on voiceprint features; a voiceprint tracing reasoning application system for fake voice is constructed based on the voiceprint features after noise reduction, and the tracing of a fake voice suspect in a cross-channel scene is realized through a speaker recognition or authentication algorithm. The application can realize accurate extraction of real speaker voiceprint features of fake voice, and through the construction of a voiceprint tracing module with confusion feature elimination and the model training design of channel noise reduction, the accuracy of voice evidence and the reliability of actual application are significantly improved.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Source speaker tracing method and device in voice conversion deep counterfeiting, and storage medium

    CN118197330A

  • Cross-channel voiceprint recognition model training method and device, equipment and storage medium

    CN118887961A