Research map / Probabilistic and scientific modeling / Molecules and proteins
Hidden in Plain Tokens: Simply Robust, Gradient-Free Watermark for Synthetic Audio
Georgios Milis, Yubin Qin, Yihan Wu, Heng Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add vocabulary clustering (cluster-level watermark) to KGW | KGW on Moshi (h=0) · Moshi conversational, identity | mean -log p-value: 8.51 -> 42.47 | Improved |
| add vocabulary clustering to KGW under noise attack | KGW on Moshi (h=0) · Moshi conversational, noise attack | mean -log p-value: 2.23 -> 20.59 | Improved |
| add vocabulary clustering to KGW under cropping | KGW on Moshi (h=0) · Moshi conversational, crop attack | mean -log p-value: 1.51 -> 16.48 | Improved |
| add vocabulary clustering versus codec finetuning (WMAR) | Moshi watermark · Moshi conversational, identity | mean -log p-value: 17.44 -> 42.47 | Improved |
| add vocabulary clustering to KGW on LibriSpeech prompts | KGW on Moshi (h=0) · Moshi LibriSpeech, identity | mean -log p-value: 3.61 -> 22.77 | Improved |
| add vocabulary clustering to KGW, audio quality | KGW on Moshi (h=0) · Moshi conversational, h=0 | NISQA MOS: 3.46 -> 3.53 | No clear effect |
| compare clustering to WMAR finetuning on audio quality | Moshi watermark · Moshi conversational, h=0 | FAD VGGish: 0.407 -> 0.133 | Improved |
| add clustering to KGW on text-to-speech | CosyVoice3 · CosyVoice3 | mean -log p-value: 1.564 -> 13.927 | Improved |
| add clustering to KGW on text-to-speech | Spark-TTS · Spark-TTS | mean -log p-value: 9.237 -> 17.806 | Improved |
| add clustering to KGW, speech recognition errors | CosyVoice3 · CosyVoice3 | WER: 0.0586 -> 0.0519 | Improved |
| add clustering to KGW, FAD VGGish | Spark-TTS · Spark-TTS | FAD VGGish: 0.2221 -> 0.3506 | Hurt |
| use Leiden instead of Louvain community detection (ablation) | vocabulary clustering · Moshi | watermark detectability | Improved |
Closest papers by meaning
- AudioMarkBench: Benchmarking Robustness of Audio Watermarking (NeurIPS 2024 · similarity 0.91)
- Robust Distortion-Free Watermark for Autoregressive Audio Generation Models (NeurIPS 2025 · similarity 0.89)
- XAttnMark: Learning Robust Audio Watermarking with Cross-Attention (ICML 2025 · similarity 0.89)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.88)
- Watermarking Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
