wavepainter

Released checkpoints for wavepainter, a masked-span speech editor.

On the English full split of Ming-Freeform-Audio-Edit, against Ren et al. (arXiv:2602.00560, the row with GRPO):

edit type	metric	Ren et al.	wavepainter (a=0.85)	wavepainter (a=0.70)
substitution	WER	4.41	3.063	3.381
insertion	WER	4.97	4.128	3.817
deletion	WER	6.88	9.136	9.525
substitution	SIM	0.78	0.943	0.943
insertion	SIM	0.82	0.961	0.960
deletion	SIM	0.78	0.918	0.918

Five of the six compared numbers improve on theirs: substitution and insertion WER, and speaker similarity on all three edit types. We do not beat them on deletion WER -- see the repository's Limitations, which also locates most of that gap with a ground-truth waveform splice.

Signal quality is reported full-reference, since a no-reference MOS estimator rates the untouched original recordings below synthesised output on this benchmark. Over the audio the edit does not touch, PESQ is 4.089 / 4.083 / 4.076 for substitution / insertion / deletion, against 4.351 for copy-synthesis with no edit at all.

Which checkpoint to use

phase2-hubert-a085 is the released model and the artifact of record. phase2-hubert-a070 is the same phase-2 child interpolated back toward the phase-1 base at a lower coefficient: it is better on insertion (3.82 vs 4.13 WER) and worse on substitution and deletion. The interpolation coefficient has no single optimum across edit types, so both are published rather than one being presented as strictly better. Speaker similarity is identical between them -- the coefficient moves only the diffusion denoiser.

Files

• phase1-base/ — the phase-1 editor. Start here to re-run phase 2 (~1 GPU-hour).
• phase2-hubert-a085/ — the released model, and the artifact of record.
• phase2-hubert-a070/ — the same child at a lower interpolation coefficient; better insertion, worse substitution and deletion.

Use

git clone https://github.com/pujariaditya/wavepainter && cd wavepainter
pip install -e . && ./setup.sh
python scripts/download_weights.py
./scripts/verify_benchmark.sh

Checkpoints are sha256-pinned in scripts/download_weights.py. Licence and attribution, including for the components these weights derive from, are in the repository's ATTRIBUTION.md.