Introduction
As digital editing tools like voice conversion and speech morphing become increasingly sophisticated, distinguishing authentic speech from tampered audio has become a significant challenge. This poses a threat to the reliability of audio as digital evidence. While traditional cryptography protects files during transmission, it offers no protection once the file is decrypted. Digital watermarking addresses this by embedding data directly into the signal.
This paper introduces a novel "semi-fragile" speech watermarking scheme based on Formant Enhancement Based Watermarking (FEBW). The goal is to create a watermark that survives standard "allowable" processing, such as audio compression, but fails when the underlying speech characteristics are maliciously altered.
System Overview
The proposed system operates across a sender and receiver side, with a transmission channel that may involve tampering.
At the sender side, a watermark $s(m)$ is embedded into the original signal $x(n)$ to produce the watermarked signal $y(n)$. After transmission, the receiver performs watermark detection to extract $\hat{s}(m)$. By comparing the original embedded watermark with the detected one, the system can generate a tampering result, localizing any unauthorized edits in the time domain.
Methodology: Formant Enhancement
The core innovation lies in manipulating Linear Spectral Frequencies (LSFs) to sharpen specific vocal resonances, known as formants. LSFs are chosen because they are stable, directly related to formants, and are standard parameters in modern speech codecs like CELP.
Bit Embedding Logic
To embed a bit, the system identifies the two sharpest formants in a 25ms frame. The embedding process follows a specific logic:
- Embedding "0": The bandwidth of the first (sharpest) formant is narrowed.
- Embedding "1": The bandwidth of the second sharpest formant is narrowed.
This narrowing is achieved by symmetrically shifting a pair of LSFs toward each other. Symmetric shifting ensures the center frequency of the formant remains unchanged, which is vital for maintaining natural sound quality and ensuring the watermark remains imperceptible.
Blind Extraction and Detection
The receiver side performs "blind" extraction, meaning the original audio is not required. The system analyzes the relative bandwidths of the formants. If the first formant is significantly sharper than the second (exceeding a predefined threshold), a "0" is extracted; otherwise, a "1" is extracted. Because standard processing (like volume scaling) affects both formants similarly, their relative relationship remains stable unless the speech structure itself is tampered with.
Experimental Results
The scheme was evaluated using the ATR database (B set), consisting of 12 Japanese sentences from 12 different speakers.
- Inaudibility: The method maintained high audio quality, achieving a Log Spectrum Distortion (LSD) of less than 1.0 dB and a PESQ score above 3.0.
- Robustness: The Bit Detection Rate (BDR) was 100% for G.711 and G.726 codecs, and approximately 90% for the G.729 codec. It also proved resilient against re-sampling and re-quantization.
- Fragility and Localization: When subjected to malicious tampering like pitch shifting or splicing, the BDR dropped significantly (e.g., to 45.19% for white noise). This allowed the system to successfully identify and localize tampered segments within a 256-second test signal.
Conclusion and Future Work
The FEBW scheme provides a stealthy and effective way to ensure speech integrity by leveraging the natural "roughness" of vocal structures. However, limitations remain regarding synchronization; if an attacker crops a segment that is not a multiple of the frame size, the watermark can desynchronize. Future work will focus on developing automatic frame synchronization schemes and improving performance under low-bitrate codecs like G.729.
Summary updated