Instructions to use committa/it_IT-serena-medium with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Piper
How to use committa/it_IT-serena-medium with Piper:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
it_IT-serena-medium β Piper voice (Italian)
Custom Italian voice for Piper, trained from the
committa/serena-synthetic-it-28h dataset (Qwen3-TTS synthetic voice).
- Language: Italian (
it) β neutral Standard Italian (no regional accent) - Voice: Female ("Serena")
- Sample rate: 22050 Hz
- Quality: medium
- Files:
it_IT-serena-medium.onnx+it_IT-serena-medium.onnx.json(must stay together)
Voice provenance
This voice model was trained using the serena-synthetic-it-28h dataset.
The dataset audio was synthetically generated using Qwen3-TTS. A short synthetic reference clip was first generated with the default Serena speaker in Qwen3-TTS CustomVoice. This clip was then used as a fixed reference with Qwen3-TTS Base in voice-cloning (ICL) mode to generate the dataset samples.
The dataset author did not record or directly use recordings of real people as reference audio. The provenance and consent status of any underlying voice recordings potentially used in Qwen3-TTS training have not been independently verified.
Recommended inference settings
The .onnx.json already embeds these defaults; the CLI uses them automatically.
They were chosen after stability testing: the Piper stock defaults (0.667 / 0.8)
tend to swallow phonemes (e.g. /r/ in "portato").
| Parameter | Value |
|---|---|
| noise_scale | 0.6 |
| length_scale | 1.0 |
| noise_w | 0.4 |
| sentence_silence | 0.2 |
Sample
CLI usage
piper -m it_IT-serena-medium.onnx < text.txt
Explicit overrides (optional):
piper -m it_IT-serena-medium.onnx --noise-scale 0.6 --noise-w 0.4 < text.txt
Home Assistant (piper add-on)
The Home Assistant add-on ignores the .onnx.json settings and always passes its
own options, whose defaults (noise_scale 0.667, noise_w 0.333) are NOT optimal for
this voice. For the best quality set in the add-on options:
noise_scale: 0.6length_scale: 1.0noise_w: 0.4sentence_silence: 0.2 (optional; add-on default is 0.0)
Notes
- Pronunciation of rare vocabulary (<10 occurrences in training data) can be off; fix with espeak-ng rules if needed.
- The training checkpoint (base for finetuning other Italian voices) is published at: https://huggingface.co/committa/it_IT-serena-medium-checkpoint
- The CC BY 4.0 license applies to the voice model files distributed in this repository. Piper and its runtime dependencies are separate software components and are subject to their respective licenses.
License & Attribution
This voice model is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
Created by Marco Baccarani
Company: Committa β https://committa.it
- Website: https://committa.it
- Hugging Face: https://huggingface.co/committa
- Source dataset: https://huggingface.co/datasets/committa/serena-synthetic-it-28h
- Dataset generation tool: https://github.com/Committa/qwen3-tts-dataset-generator
Preferred attribution
If you use or redistribute this voice model, please credit:
it_IT-serena-medium by Marco Baccarani.
Company: Committa β https://committa.it
Source: https://huggingface.co/committa/it_IT-serena-medium
License: CC BY 4.0.
- Downloads last month
- 190