DuraMark embeds watermarks in AI speech via syllable duration editing
New watermarking method survives neural codecs and vocoders by hiding data in timing.
Researchers Zhenwei Mou and colleagues have proposed DuraMark, a novel watermarking framework designed specifically for LLM-based text-to-speech (TTS) systems. Unlike conventional signal-level methods that embed watermarks into waveforms or spectrograms—rendering them vulnerable to generative attacks such as neural codecs and vocoders—DuraMark operates at an information level by manipulating syllable durations. The framework integrates a duration-controllable LLM-based TTS model that edits syllable timing during the speech synthesis process, effectively hiding a traceable watermark within the normal variations of speech rhythm.
DuraMark's detection pipeline uses a dedicated duration extractor to retrieve the embedded watermarks from the synthesized audio. In experiments, DuraMark significantly outperformed signal-level baselines, demonstrating robust watermark retention even after the speech was passed through generative post-processing steps. This resilience is critical as AI voice cloning becomes more prevalent and deepfake detection becomes harder. The paper, accepted at INTERSPEECH 2026, includes audio samples on a public webpage. By embedding watermarks in a domain resistant to compression and generative transformation, DuraMark offers a practical tool for authenticating AI-generated speech.
- DuraMark edits syllable durations during LLM-based TTS synthesis to embed watermarks, avoiding vulnerable signal-level methods.
- Achieves superior robustness against generative attacks like neural codecs and vocoders compared to traditional waveform/spectrogram watermarking.
- Accepted to INTERSPEECH 2026; audio samples are publicly available for validation.
Why It Matters
A robust watermarking method for AI voice cloning helps combat deepfakes and trace generated speech.