Update README.md
Browse filesremoved meaningless comments
README.md
CHANGED
|
@@ -24,7 +24,7 @@ Custom Italian voice for [Piper](https://github.com/OHF-Voice/piper1-gpl), train
|
|
| 24 |
- Quality: medium
|
| 25 |
- Files: `it_IT-serena-medium.onnx` + `it_IT-serena-medium.onnx.json` (must stay together)
|
| 26 |
|
| 27 |
-
## v2 — fine-tuned
|
| 28 |
|
| 29 |
This release fixes a defect of the previous model: words at the start of a
|
| 30 |
sentence after `.?!` (e.g. "Purtroppo. **Domani** non posso.") were spoken
|
|
@@ -33,14 +33,6 @@ with a collapsed, almost inaudible duration.
|
|
| 33 |
Fix: fine-tuned from the v1 checkpoint on an extended dataset
|
| 34 |
(`serena-synthetic-it-28h`, +1,372 targeted sentence-initial phrases).
|
| 35 |
|
| 36 |
-
Validation result (71-sentence regression scan, duration ratio vs healthy
|
| 37 |
-
median of the same word):
|
| 38 |
-
|
| 39 |
-
| run | collapsed words |
|
| 40 |
-
|-----|-----------------|
|
| 41 |
-
| v1 | 4/8 |
|
| 42 |
-
| **v2 (this)** | **2/41** (remaining 2 are analysis artifacts; "domani" is correctly voiced) |
|
| 43 |
-
|
| 44 |
Inference defaults are unchanged (noise_scale 0.6 / noise_w 0.4), so existing
|
| 45 |
users get the fix without changing any setting.
|
| 46 |
|
|
|
|
| 24 |
- Quality: medium
|
| 25 |
- Files: `it_IT-serena-medium.onnx` + `it_IT-serena-medium.onnx.json` (must stay together)
|
| 26 |
|
| 27 |
+
## v2 — fine-tuned
|
| 28 |
|
| 29 |
This release fixes a defect of the previous model: words at the start of a
|
| 30 |
sentence after `.?!` (e.g. "Purtroppo. **Domani** non posso.") were spoken
|
|
|
|
| 33 |
Fix: fine-tuned from the v1 checkpoint on an extended dataset
|
| 34 |
(`serena-synthetic-it-28h`, +1,372 targeted sentence-initial phrases).
|
| 35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
Inference defaults are unchanged (noise_scale 0.6 / noise_w 0.4), so existing
|
| 37 |
users get the fix without changing any setting.
|
| 38 |
|