rhasspy

/

piper-voices

like 652

Follow

Rhasspy 321

What this PR changes

Updates it/it_IT/serena/{medium,high}/ in the voice catalog with the fixed version
of both it_IT-serena-medium and it_IT-serena-high voices:

• it_IT-serena-{medium,high}.onnx — re-exported from the fine-tuned checkpoints (v2)
• it_IT-serena-{medium,high}.onnx.json — updated to match the new models (phoneme_id_map,
  piper_version 1.5.0, language block it_IT; inference defaults unchanged)
• MODEL_CARD — dataset updated 27h → 28h
• samples/speaker_0.mp3 — regenerated samples from the fixed models

Why

The previously published it_IT-serena-{medium,high} had a training-data defect: words at
the start of a sentence after .?! were spoken with collapsed, nearly inaudible
durations (e.g. "Purtroppo. Domani non posso." → "domani" barely audible, ~120 ms vs
~1 s healthy).

Fix: fine-tuned the v1 checkpoints on an extended dataset
(committa/serena-synthetic-it-28h, 30,842 clips, 28 h — the 27 h base plus 1,372
targeted sentence-initial phrases covering the missing pattern).

Validation

71-sentence regression scan (duration of sentence-initial words vs healthy median of the
same word, ratio < 0.5 = collapse):

The remaining flags in v2 are analysis artifacts (all are "domani" after ., verified by
ear as correctly pronounced — the audio splitter in the scan mis-segments that word).
No new collapses were introduced: previously-affected words ("come", "il", "sei",
"subito", "posso") are all healthy (ratio ≥ 0.88).

Inference defaults are unchanged (noise_scale 0.6 / length_scale 1.0 / noise_w 0.4),
so existing users receive the fix with no configuration change.

Compatibility notes

• The .onnx.json phoneme_id_map changed (157 → 166 entries): it now reflects the
  piper1-gpl phoneme set of the new models. The pair (onnx + json) must be updated
  together, as done here.
• Same voice ids (it_IT-serena-medium, it_IT-serena-high), same language/quality/
  sample-rate: no client-side changes required.

Related

• Dataset: https://huggingface.co/datasets/committa/serena-synthetic-it-28h
• Model repos: https://huggingface.co/committa/it_IT-serena-medium,
  https://huggingface.co/committa/it_IT-serena-high
• Checkpoints: https://huggingface.co/committa/it_IT-serena-medium-checkpoint
  (epoch=14-step=83250.ckpt), https://huggingface.co/committa/it_IT-serena-high-checkpoint
  (epoch=18-step=105450.ckpt)