Replace checkpoint with fine-tuned v2 (epoch=14-step=83250)

- Removed epoch=94-step=559930.ckpt (v1, sentence-initial word collapse defect)
- Added epoch=14-step=83250.ckpt (fine-tuned on serena-synthetic-it-28h)
- val_mel 0.3763 (best)

README.md

@@ -10,27 +10,29 @@ tags:
 - checkpoint
 - finetuning
 datasets:
-- committa/serena-synthetic-it-27h
+- committa/serena-synthetic-it-28h
 ---
 
 # it_IT-serena — Piper training checkpoint (Italian, medium)
 
-Finetuning base checkpoint for Piper TTS (VITS), trained **from scratch** on Italian.
+Finetuning base checkpoint for Piper TTS (VITS), fine-tuned on Italian.
 
 - Model: piper (piper_train), quality **medium**, 22.05 kHz
 - Base voice: **serena** — female, neutral Standard Italian (no regional accent), synthetic
-- Training: **94 epochs** from scratch, batch 10, 29,470 clips (27 h)
-- Dataset: [committa/serena-synthetic-it-27h](https://huggingface.co/datasets/committa/serena-synthetic-it-27h)
-- File: `epoch=94-step=559930.ckpt` — **keep this filename** (piper_train convention:
-  `epoch=N-step=M`), which tooling uses to pick the highest checkpoint
+- Training: **14 epochs of fine-tuning** (resumed from the v1 checkpoint, ~94 epochs
+  from scratch) on `committa/serena-synthetic-it-28h` (30,842 clips, 28 h)
+- Dataset: [committa/serena-synthetic-it-28h](https://huggingface.co/datasets/committa/serena-synthetic-it-28h)
+- File: `epoch=14-step=83250.ckpt` — piper_train convention: `epoch=N-step=M`,
+  which tooling uses to pick the highest checkpoint
+- Best validation: val_mel **0.3763** (epoch 14)
 
 ## Why this checkpoint
 
-Piper publishes **no official pretrained checkpoint for Italian** (the `rhasspy/piper-checkpoints`
-catalog has no `it` entry; the only official Italian voices are `it_IT-paola-medium`, which was
-finetuned from a U.S. English base, and `it_IT-riccardo-x_low`, trained from scratch at x_low
-quality — neither exposes a trainable checkpoint). This is the only public Italian **medium**
-checkpoint that can be used as a finetuning base.
+This is the **fixed** version of the published Italian medium checkpoint: the
+previous release had a defect (sentence-initial words after `.?!` spoken with
+collapsed duration). This checkpoint is the fine-tuned base used to build the
+fixed `it_IT-serena-medium` voice, and can be used as a base for further
+finetuning of other Italian voices.
 
 ## Sample
 
@@ -45,7 +47,7 @@ Any checkpoint of the same quality config (medium) can be resumed with piper_tra
 ```sh
 python -m piper_train \
     --dataset-dir /path/to/your_voice_training_folder \
-    --resume_from_checkpoint epoch=94-step=559930.ckpt \
+    --resume_from_checkpoint epoch=14-step=83250.ckpt \
     --quality medium
 ```
 
@@ -56,5 +58,5 @@ enough epochs the new voice fully takes over.
 
 ## License
 
-CC-BY-4.0 — derived from [committa/serena-synthetic-it-27h](https://huggingface.co/datasets/committa/serena-synthetic-it-27h).
+CC-BY-4.0 — derived from [committa/serena-synthetic-it-28h](https://huggingface.co/datasets/committa/serena-synthetic-it-28h).
 If you use this checkpoint to build a voice, credit the dataset and this model.

epoch=94-step=559930.ckpt → epoch=14-step=83250.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3e20efaaf795520b01e4fca6088d58040cd27da28f30e6fab0778d2b9d27d781
-size 845898456
+oid sha256:e05a83db9b226d2e5ee3e5b314e36c937b5bfddfa99ee2ba63305a88aa4b7622
+size 845892045