DuraMark: Duration-Embedded Watermarking in LLM-based TTS

Authors: Zhenwei Mou, Weili Jiang, Liping Chen, Zhen-Hua Ling, Kong Aik Lee, Kai Gao, Boyu Zhao

Published: 2026-06-13 11:53:06+00:00

Comment: Accepted to INTERSPEECH 2026. 5 pages, 1 figure. Audio samples: https://muzw.github.io/duramark_demo/

AI Summary

This paper introduces DuraMark, a robust information-level watermarking framework designed for LLM-based Text-to-Speech (TTS) models to mitigate deepfake misuse. Unlike signal-level methods vulnerable to generative attacks, DuraMark embeds watermarks by editing syllable durations during speech synthesis. It demonstrates superior robustness against various generative attacks while maintaining high speech naturalness.

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/.

Key findings

DuraMark achieved superior robustness against generative attacks, such as neural audio codecs and vocoders, consistently maintaining a True Positive Rate (TPR) above 95%, significantly outperforming signal-level baselines. It also demonstrated robust performance against traditional signal processing, lossy compression, and speech enhancement. The method showed minimal impact on speech naturalness, effectively preserving high speech quality as measured by Character Error Rate (CER) and Mean Opinion Score (MOS).

Approach

DuraMark embeds watermarks by integrating a duration-controllable LLM-based TTS model that precisely edits syllable durations during speech synthesis. Watermark bits dictate whether a syllable's duration should be even or odd, modifying the predicted duration accordingly. For detection, a duration extractor recovers the syllable duration sequence from the speech, and a correlation score between the extracted durations and the watermark sequence determines its presence.

Datasets

WenetSpeech (training), AISHELL-3 (evaluation)

Model(s)

Duration-controllable LLM-based TTS (based on CosyVoice framework, integrating an LLM with a flow matching decoder), Transformer-based Duration Extractor, Optimal Transport Conditional Flow Matching (OT-CFM) decoder.

Author countries

China

← Previous