DuraMark: Duration-Embedded Watermarking in LLM-based TTS

2026-06-13

AI Summary

This work proposes DuraMark, a novel information-layer watermarking framework that addresses the vulnerability of existing signal-level audio watermarking methods to generative attacks such as neural codecs. Unlike prior approaches that embed watermarks directly into the acoustic signal, DuraMark elevates watermark insertion to the linguistic information layer by modulating syllable durations during LLM-driven speech synthesis. A dedicated duration extractor is employed to robustly detect the embedded watermark from the synthesized speech. Experimental results demonstrate that DuraMark significantly outperforms state-of-the-art signal-level methods in watermark robustness against neural vocoders and codec-based transformations, offering a promising direction for resilient audio watermarking in the era of generative speech models.

Abstract

Large language model (LLM)-based text-to-speech (TTS) models have achieved remarkable voice cloning capabilities, raising concerns about potential deepfake misuse. Speech watermarking mitigates this by embedding traceable information into generated speech. Mainstream watermarking methods operate at the signal level (waveform or spectrogram), rendering the watermark vulnerable to generative attacks (e.g., neural codec and vocoder). To address this, we propose DuraMark, a robust information-level watermarking framework. It utilizes syllable duration editing to achieve watermark embedding. Specifically, DuraMark integrates a duration-controllable LLM-based TTS model to edit syllable durations during synthesis, coupled with a duration extractor to extract these durations for detection. Experiments demonstrate DuraMark's superior robustness against generative attacks, significantly outperforming signal-level baselines. Audio samples are available at https://muzw.github.io/duramark_demo/.

Problem

• speech watermarking
• deepfake
• LLM-based TTS
• generative attacks
• voice cloning

Innovation

• watermarking
• duration editing
• LLM-based TTS
• generative attack robustness
• information-level embedding