committa/serena-synthetic-it-27h

Qwen3-TTS Italian Synthetic Speech (27h)

Synthetic Italian single-speaker speech dataset for TTS training (e.g. Piper), generated with Qwen3-TTS-1.7B-Base in voice-cloning mode. ~29.5k clips, ~27 hours, 22.05 kHz mono WAV, Piper-ready metadata.

Dataset summary

Property	Value
Clips (train / val)	26,523 / 2,947
Total duration	~27.3 h (98,099 s)
Sample rate	22,050 Hz mono, 16-bit WAV
Loudness	Normalized to -23 LUFS, silence-trimmed
Language	Italian
Speakers	1 (synthetic voice, see below)
Metadata format	`wavs/000000-008999/NNNNNN.wav

Generation & provenance

• Tool: generated with qwen3-tts-dataset-generator, an end-to-end pipeline (generation → validation → pronunciation check → normalization → publish) for synthetic TTS datasets.
• Model: Qwen3-TTS-1.7B-Base (Qwen/Qwen3-TTS-12Hz-1.7b-Base), bfloat16, voice cloning via ICL (in-context learning) from a fixed reference clip, x-vector-only disabled.
• Voice: a single fixed synthetic reference voice, kept constant across all clips for maximum cross-clip consistency (stable voice workflow).
• Instruct: "Read in a calm, neutral tone."
• Sampling: temperature=0.4, top_k=50, top_p=1.0, repetition_penalty=1.05, max_new_tokens=512, seed=42, do_sample=true.

Corpus composition & attribution

The sentence corpus (29,475 sentences) is a mix of two sources:

• ~90% Tatoeba-derived sentences (Italian, CC-BY 2.0 FR, from the official export at https://downloads.tatoeba.org/exports/per_language/ita/ita_sentences.tsv.bz2), adapted: English/American given names were replaced with Italian names, and the text received editorial corrections and accent fixes. They are therefore derivative/adapted works, not verbatim copies.
• ~10% original LLM-authored sentences across conversational, technical and literary registers, written specifically for this corpus (models: Claude Sonnet, GPT-4o, Qwen, Gemini, DeepSeek; one production batch was excluded after a quality review).

The corpus was assembled with deterministic seed=42 shuffles; the name-replacement and LLM batch verification reports are included in the qwen3-tts-dataset-generator project.

Quality & validation

Each clip passed a two-stage automated validation:

1. Word-level: ASR (faster-whisper large-v3) word error rate must be ≤ 0.25.
2. Phoneme-level: recognition to espeak-ng phonemes (facebook/wav2vec2-xlsr-53-espeak-cv-ft) compared against the phonemized reference; Phoneme Error Rate must be ≤ 0.30.

Measured phoneme PER distribution over the final corpus (n = 29,475):

Stat	PER
mean	0.168
median	0.167
p75	0.208
p90	0.247

File structure

README.md
LICENSE              CC-BY-4.0 (see also attribution section above)
report.json          Generation/report metadata (model, sampling, stats)
metadata_train.csv   | path|transcript (26,523 rows)
metadata_val.csv     | path|transcript (2,947 rows)
wavs/000000-008999/  | 8,996 clips
wavs/009000-017999/  | 9,000 clips
wavs/018000-026999/  | 8,999 clips
wavs/027000-029474/  | 2,475 clips

The wavs are spread over subdirectories of at most 9,000 files each so the dataset can be committed to the Hugging Face Hub, which rejects commits with more than 10,000 files in a single directory.

metadata_*.csv paths are relative to this repository root, so the folder can be used directly as a Piper dataset directory.

Usage

Piper

The metadata format is Piper-native (wavs/NNNNNN.wav|transcript). For training, concatenate the split files into a single metadata.csv and pass the repository root as --dataset-dir. See the Piper docs for training details.

Python / datasets

The layout is plain files, not AudioFolder, so load it manually:

import pandas as pd
from datasets import Dataset, Audio

df = pd.read_csv("hf://datasets/committa/serena-synthetic-it-27h/metadata_train.csv",
                 sep="|", names=["audio", "text"])
ds = Dataset.from_pandas(df).cast_column("audio", Audio(sampling_rate=22050))

Audio samples

Sample 1:

Sample 2:

Sample 3:

Ethical considerations

All audio is synthetic: it is the output of a text-to-speech model and does not reproduce the voice of any real person. No consent issues apply, but be aware that the recordings could be used to build voice-cloning tools, so treat the dataset accordingly.

License

The dataset is released under CC-BY-4.0 (see LICENSE). Attribution:

• Audio: generated by the pipeline, © the dataset author.
• ~90% of transcripts: adapted from Tatoeba sentences (CC-BY 2.0 FR, https://tatoeba.org), modified (name substitution, corrections, accents).
• ~10% of transcripts: original LLM-authored sentences, © the dataset author.