Add Italian it_IT-serena medium and high checkpoints
New Italian checkpoint for the catalog — the first one for it (piper-checkpoints
currently has no Italian entry).
- Language: it_IT (Italian, Italy)
- Quality: medium, 22,050 Hz
- Voice: serena — female, neutral Standard Italian (no regional accent), synthetic
- Training: trained from scratch (94 epochs, batch 10, ~27 h of audio)
- Dataset: committa/serena-synthetic-it-27h (CC-BY-4.0)
- Recommended inference settings (embedded in the config):
noise_scale 0.6 / length_scale 1.0 / noise_w 0.4
Files follow the existing piper-checkpoints layout
(<family>/<lang_code>/<dataset>/<quality>/, cf. en/en_US/lessac/medium/):epoch=94-step=559930.ckpt, config.json, dataset.jsonl.gz, MODEL_CARD,train.sh, .gitattributes.
Why this checkpoint
There is no official pretrained checkpoint for Italian (the catalog has no it
entry), and the only official Italian voices expose no trainable checkpoint
(it_IT-paola-medium was finetuned from U.S. English and only the onnx is
published; it_IT-riccardo-x_low is x_low). This checkpoint is a valid medium
base for finetuning other Italian voices — female or male targets both work,
and the model already speaks Italian (phonemes and prosody), so new voices
converge much faster than from scratch.
Repositories
- Checkpoint: https://huggingface.co/committa/it_IT-serena-medium-checkpoint
- Exported voice (onnx): https://huggingface.co/committa/it_IT-serena-medium
(also submitted to piper-voices, PR #95)
License
CC-BY-4.0 — derived from
committa/serena-synthetic-it-27h.
New Italian checkpoints for the catalog - the first ones for it
(piper-checkpoints currently has no Italian entry at all).
- Language: it_IT (Italian, Italy)
- Quality: medium (94 epochs) and high (99 epochs), 22,050 Hz
- Voice: serena - female, neutral Standard Italian (no regional accent), synthetic
- Training: trained from scratch (batch 10, ~27 h of audio)
- Dataset: committa/serena-synthetic-it-27h (CC-BY-4.0)
- Recommended inference settings (embedded in the config):
noise_scale 0.6 / length_scale 1.0 / noise_w 0.4
Files follow the existing piper-checkpoints layout
(////, cf. en/en_US/lessac/medium/):
it/it_IT/serena/medium/ (epoch=94-step=559930.ckpt, config.json,
dataset.jsonl.gz, MODEL_CARD, train.sh, .gitattributes)
it/it_IT/serena/high/ (epoch=99-step=604140.ckpt, config.json,
dataset.jsonl.gz, MODEL_CARD, train.sh, .gitattributes)
Why this checkpoint
There is no official pretrained checkpoint for Italian (the catalog has no it
entry), and the only official Italian voices expose no trainable checkpoint
(it_IT-paola-medium was finetuned from U.S. English and only the onnx is
published; it_IT-riccardo-x_low is x_low). These checkpoints are valid
medium/high bases for finetuning other Italian voices - female or male
targets both work, and the models already speak Italian (phonemes and
prosody), so new voices converge much faster than from scratch.
Repositories
- Checkpoints:
medium: https://huggingface.co/committa/it_IT-serena-medium-checkpoint
high: https://huggingface.co/committa/it_IT-serena-high-checkpoint - Exported voices (onnx):
medium: https://huggingface.co/committa/it_IT-serena-medium
high: https://huggingface.co/committa/it_IT-serena-high
(also submitted to piper-voices, PR "Add it_IT-serena-medium and
it_IT-serena-high (Italian)")
License
CC-BY-4.0 - derived from committa/serena-synthetic-it-27h.