Watermarking Based on Compressive Sensing for Digital Speech Detection and Recovery ^†

Area of Science:

• Digital Signal Processing
• Multimedia Security
• Speech Processing

Background:

• Digital audio content is vulnerable to malicious tampering.
• Existing methods for speech tampering detection and recovery often lack imperceptibility or robustness.
• A robust and imperceptible solution is needed for verifying speech integrity.

Purpose of the Study:

• To propose a novel imperceptible, fragile, and blind watermark scheme for speech tampering detection and self-recovery.
• To accurately localize tampered regions within a speech signal.
• To recover the original speech content from tampered segments.

Main Methods:

• Embedding watermark data derived from Discrete Cosine Transform (DCT) coefficients.
• Sharing watermark information across frame groups to balance data waste and tampering coincidence.
• Utilizing compressive sensing techniques exploiting DCT-domain sparseness for coefficient retrieval.
• Implementing a deep learning-based enhancement for improving recovered speech Signal-to-Noise Ratio (SNR).

Main Results:

• The proposed watermarking scheme is imperceptible.
• Accurate localization of tampered speech areas is achieved.
• Recovered speech remains intelligible at high tampering rates (up to 47.6%).
• Deep learning enhancement significantly improves the SNR of recovered speech.

Conclusions:

• The developed watermark scheme offers effective speech tampering detection and self-recovery.
• The method provides a good trade-off between data integrity and recovery quality.
• The integration of compressive sensing and deep learning enhances the practical applicability for multimedia security.