Datasets:

---

rhasspy

/

piper-checkpoints

like 78

Follow

Rhasspy 319

What this PR changes

Adds the first Italian entries to the rhasspy/piper-checkpoints catalog:

• it/it_IT/serena/medium/ — checkpoint epoch=14-step=83250.ckpt
• it/it_IT/serena/high/ — checkpoint epoch=18-step=105450.ckpt

Each folder follows the existing piper-checkpoints layout
(<family>/<lang_code>/<dataset>/<quality>/, cf. en/en_US/lessac/medium/):
epoch=N-step=M.ckpt, config.json, dataset.jsonl.gz, MODEL_CARD,
train.sh.

Why this checkpoint

The rhasspy/piper-checkpoints catalog has no Italian entry — piper publishes no
official pretrained checkpoint for Italian. The only official Italian voices are
it_IT-paola-medium (finetuned from a U.S. English base, no trainable checkpoint
published) and it_IT-riccardo-x_low (x_low quality, from scratch, no checkpoint
exposed). These are the first trainable medium and high Italian checkpoints
in the catalog.

• Language: it_IT (Italian, Italy)
• Quality: medium and high, 22,050 Hz
• Voice: serena — female, neutral Standard Italian (no regional accent), synthetic
• Training: trained from scratch on the 27 h base, then fine-tuned on the extended
  serena-synthetic-it-28h dataset: medium 14 epochs (val_mel 0.3763),
  high 18 epochs (val_mel 0.3567), batch 10
• Dataset: https://huggingface.co/datasets/committa/serena-synthetic-it-28h (CC-BY-4.0)

Repositories

• Checkpoints: https://huggingface.co/committa/it_IT-serena-medium-checkpoint,
  https://huggingface.co/committa/it_IT-serena-high-checkpoint
• Exported voices (onnx): https://huggingface.co/committa/it_IT-serena-medium,
  https://huggingface.co/committa/it_IT-serena-high
  (also in piper-voices, same voice ids)

License

CC-BY-4.0 — derived from committa/serena-synthetic-it-28h.