Synthetic Italian voice

Serena

An Italian female voice for Piper, trained on a 28-hour synthetic dataset. Two versions, same voice: medium for resource-constrained devices, high for maximum quality.

Download Medium Download High Dataset on Hugging Face
100% synthetic voice
CC BY 4.0
Ready for Piper and Home Assistant

Listen to the voice

The same sample, generated by both versions of the model.

it_IT-serena-medium · Medium

Audio sample generated by the distributed model.

it_IT-serena-high · High

Audio sample generated by the distributed model.

Voice provenance

Serena is an entirely synthetic voice. It was not recorded from a person: a short reference clip was generated with the default “Serena” speaker in Qwen3-TTS CustomVoice, then used as a fixed reference with Qwen3-TTS-1.7B-Base in voice-cloning (ICL) mode to produce all 30,842 clips in the dataset.

The same reference clip was reused throughout the generation, keeping timbre and pitch consistent. The instruction given to the model was “Read in a calm, neutral tone.”

The dataset

The model is trained on the serena-synthetic-it-28h dataset, released under CC BY 4.0.

  • Clips: 30,842 (27,895 train / 2,947 validation)
  • Duration: 28h 16min (28.27 hours)
  • Format: WAV mono, 22,050 Hz, 16 bit
  • Loudness: normalized to -23 LUFS, silence-trimmed
  • Language: neutral standard Italian, no regional accent
  • ASR validation: WER ≤ 0.25 with faster-whisper large-v3
  • Phoneme validation: PER ≤ 0.30 (wav2vec2 CTC + espeak-ng), mean PER 0.168
  • Metadata: LJSpeech format (path|transcript), ready for Piper training

Corpus composition

The 30,842 sentences come from two sources:

  • ~90% Tatoeba-derived sentences (Italian, CC BY 2.0 FR), adapted: English/American given names were replaced with Italian names and the text received editorial corrections and accent fixes. They are therefore derivative works, not verbatim copies. No Tatoeba audio was used.
  • ~10% original LLM-authored sentences (Claude Sonnet, GPT-4o, Qwen, Gemini, DeepSeek), written specifically for the corpus across conversational, technical and literary registers.

Technical specs

Language
Neutral standard Italian, no regional accent
Gender
Female voice
Sample rate
22,050 Hz, mono, 16 bit
Format
Piper / ONNX (.onnx + .onnx.json)
License
CC BY 4.0
Dataset
28.27 hours, 30,842 clips

Medium or high?

Both versions share the same voice and the same inference settings: what changes is model size and perceived quality.

ModelQualitySizeRecommended use
it_IT-serena-mediumMedium60.6 MBResource-constrained devices, real-time synthesis, Home Assistant
it_IT-serena-highHigh108.9 MBMaximum quality for narration and audiobooks; needs more CPU and RAM

How Serena was made

The pipeline that turns a text corpus into a validated TTS dataset, released as an open-source project.

1

Generation

Qwen3-TTS-1.7B-Base in voice-cloning (ICL) mode generates the audio from the text corpus, on GPU, with a resumable checkpoint.

2

ASR validation

faster-whisper large-v3 transcribes every clip and rejects those with WER above 0.25.

3

Phoneme check

wav2vec2 + espeak-ng compare phonemes and catch pronunciation errors that ASR misses (PER ≤ 0.30).

4

Normalization

Mono, 22,050 Hz, loudness to -23 LUFS, silence trimmed, 16-bit PCM.

5

Publishing

LJSpeech manifest, deterministic train/val split and a Hugging Face-ready archive.

Generator source code on GitHub

Usage notes

The .onnx.json files already embed the recommended settings, chosen after stability testing: Piper’s stock defaults (0.667 / 0.8) tend to swallow some phonemes, for example the /r/ in “portato”.

  • noise_scale: 0.6
  • length_scale: 1.0
  • noise_w: 0.4
  • sentence_silence: 0.2

The Home Assistant Piper add-on ignores the .onnx.json values and always passes its own: for the best quality it must be configured manually with the same values.

Pronunciation of rare vocabulary (fewer than 10 occurrences in the training set) can be off; it can be fixed with espeak-ng rules. The training checkpoints are published as a base for fine-tuning other Italian voices: serena-high-checkpoint and serena-medium-checkpoint.

Limitations and ethical considerations

Like any synthetic voice, Serena can be used to train voice-cloning systems. Users should consider the potential for misuse and make sure their applications comply with applicable laws and ethical guidelines.

Licenses and terms of use

Each artifact has its own license: model and dataset are CC BY 4.0 and require attribution; the generator code is MIT; third-party components keep their own terms.

ArtifactLicenseNotes
Piper models (medium and high)CC BY 4.0Attribution required. .onnx and .onnx.json must be distributed together.
serena-synthetic-it-28h datasetCC BY 4.030,842 clips, 28.27 hours, with dataset card and audio samples.
qwen3-tts-dataset-generatorMITCode of the generation and validation pipeline.
Qwen3-TTSApache-2.0Model used to generate the dataset audio.
Piper (piper1-gpl)GPL-3.0Separate runtime: not covered by the model license.
Tatoeba sentences (adapted)CC BY 2.0 FR~90% of transcripts. No Tatoeba audio was used.

Preferred attribution

Use the block for the version you distribute. The dataset has its own attribution text, stated in the dataset card.

Version medium

it_IT-serena-medium by Marco Baccarani.
Company: Committa — https://committa.it
Source: https://huggingface.co/committa/it_IT-serena-medium
License: CC BY 4.0.

Version high

it_IT-serena-high by Marco Baccarani.
Company: Committa — https://committa.it
Source: https://huggingface.co/committa/it_IT-serena-high
License: CC BY 4.0.

Same text as the model card: if you redistribute the model, keep the name, URL and license.

Need an Italian voice for your project?

We can train, evaluate and integrate custom Piper voices, on-premise or inside a product.

Contact us