---
library_name: piper
license: cc-by-4.0
language:
- it
tags:
- text-to-speech
- tts
- piper
- checkpoint
- finetuning
datasets:
- committa/serena-synthetic-it-28h
---

# it_IT-serena — Piper training checkpoint (Italian, medium)

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: **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. It is the base used to build
`it_IT-serena-medium`, and it is fine-tuned on an extended dataset
(`serena-synthetic-it-28h`) with targeted sentence-initial phrases for improved
prosody at the start of sentences.

## Sample

Audio of the base voice (it_IT-serena-medium):

<audio controls src="https://huggingface.co/committa/it_IT-serena-medium-checkpoint/resolve/main/samples/speaker_0.mp3"></audio>

## Use as a base for finetuning other Italian voices

Any checkpoint of the same quality config (medium) can be resumed with piper_train:

```sh
python -m piper_train \
    --dataset-dir /path/to/your_voice_training_folder \
    --resume_from_checkpoint epoch=14-step=83250.ckpt \
    --quality medium
```

Because the base already speaks Italian (correct phonemes and prosody), a new Italian voice
converges in **far fewer epochs** than from scratch — and with less data. This works for
**both female and male target voices**: the timbre is learned from the new dataset, so with
enough epochs the new voice fully takes over.

## License

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.