Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -15,26 +15,38 @@ a masked-span speech editor.
|
|
| 15 |
On the English `full` split of Ming-Freeform-Audio-Edit, against Ren et al.
|
| 16 |
([arXiv:2602.00560](https://arxiv.org/abs/2602.00560), the row with GRPO):
|
| 17 |
|
| 18 |
-
| edit type | metric | Ren et al. | wavepainter |
|
| 19 |
-
|---|---|---|---|
|
| 20 |
-
| substitution | WER | 4.41 | **3.063** |
|
| 21 |
-
| insertion | WER | 4.97 |
|
| 22 |
-
| deletion | WER | **6.88** | 9.136 |
|
| 23 |
-
| substitution | SIM | 0.78 | **0.943** |
|
| 24 |
-
| insertion | SIM | 0.82 | **0.961** |
|
| 25 |
-
| deletion | SIM | 0.78 | **0.918** |
|
| 26 |
-
| substitution | DNSMOS | **3.11** | 2.931 |
|
| 27 |
-
| insertion | DNSMOS | **3.18** | 2.905 |
|
| 28 |
-
| deletion | DNSMOS | **3.09** | 2.938 |
|
| 29 |
|
| 30 |
Five of the nine numbers they report improve on theirs: substitution and
|
| 31 |
insertion WER, and speaker similarity on all three edit types. We do not beat
|
| 32 |
them on deletion WER or on DNSMOS anywhere -- see the repository's Limitations.
|
| 33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
## Files
|
| 35 |
|
| 36 |
- `phase1-base/` — the phase-1 editor. Start here to re-run phase 2 (~1 GPU-hour).
|
| 37 |
-
- `phase2-hubert-a085/` — the released model
|
|
|
|
|
|
|
| 38 |
|
| 39 |
## Use
|
| 40 |
|
|
|
|
| 15 |
On the English `full` split of Ming-Freeform-Audio-Edit, against Ren et al.
|
| 16 |
([arXiv:2602.00560](https://arxiv.org/abs/2602.00560), the row with GRPO):
|
| 17 |
|
| 18 |
+
| edit type | metric | Ren et al. | wavepainter (a=0.85) | wavepainter (a=0.70) |
|
| 19 |
+
|---|---|---|---|---|
|
| 20 |
+
| substitution | WER | 4.41 | **3.063** | 3.381 |
|
| 21 |
+
| insertion | WER | 4.97 | 4.128 | **3.817** |
|
| 22 |
+
| deletion | WER | **6.88** | 9.136 | 9.525 |
|
| 23 |
+
| substitution | SIM | 0.78 | **0.943** | **0.943** |
|
| 24 |
+
| insertion | SIM | 0.82 | **0.961** | 0.960 |
|
| 25 |
+
| deletion | SIM | 0.78 | **0.918** | **0.918** |
|
| 26 |
+
| substitution | DNSMOS | **3.11** | 2.931 | 2.925 |
|
| 27 |
+
| insertion | DNSMOS | **3.18** | 2.905 | 2.908 |
|
| 28 |
+
| deletion | DNSMOS | **3.09** | 2.938 | 2.935 |
|
| 29 |
|
| 30 |
Five of the nine numbers they report improve on theirs: substitution and
|
| 31 |
insertion WER, and speaker similarity on all three edit types. We do not beat
|
| 32 |
them on deletion WER or on DNSMOS anywhere -- see the repository's Limitations.
|
| 33 |
|
| 34 |
+
### Which checkpoint to use
|
| 35 |
+
|
| 36 |
+
`phase2-hubert-a085` is the released model and the artifact of record.
|
| 37 |
+
`phase2-hubert-a070` is the same phase-2 child interpolated back toward the
|
| 38 |
+
phase-1 base at a lower coefficient: it is better on insertion (3.82 vs 4.13
|
| 39 |
+
WER) and worse on substitution and deletion. The interpolation coefficient has
|
| 40 |
+
no single optimum across edit types, so both are published rather than one
|
| 41 |
+
being presented as strictly better. Speaker similarity is identical between
|
| 42 |
+
them -- the coefficient moves only the diffusion denoiser.
|
| 43 |
+
|
| 44 |
## Files
|
| 45 |
|
| 46 |
- `phase1-base/` — the phase-1 editor. Start here to re-run phase 2 (~1 GPU-hour).
|
| 47 |
+
- `phase2-hubert-a085/` — the released model, and the artifact of record.
|
| 48 |
+
- `phase2-hubert-a070/` — the same child at a lower interpolation coefficient;
|
| 49 |
+
better insertion, worse substitution and deletion.
|
| 50 |
|
| 51 |
## Use
|
| 52 |
|