ParsVoice — Persian Zero-Shot TTS

Paper

ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

Mohammad Javad Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery

University of Tehran

Read the paper Access the dataset Hear the samples ↓

@article{ranjbar2025parsvoice,
  title={ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis},
  author={Ranjbar Kalahroodi, Mohammad Javad and Faili, Heshaam and Shakery, Azadeh},
  journal={arXiv preprint arXiv:2510.10774},
  year={2025}
}

About ParsVoice

Existing Persian speech datasets are typically far smaller than their English counterparts, which limits Persian speech technology. ParsVoice addresses this with an automated pipeline that turns raw audiobook recordings into TTS-ready data — a BERT-based sentence-completion detector, a binary-search boundary-optimization step for precise audio-text alignment, and audio-text quality assessment tailored to Persian.

The pipeline processes 2,000 audiobooks into 3,526 hours of clean speech, filtered down to a 1,804-hour high-quality subset spanning 470+ speakers. Fine-tuning XTTS on this data reaches a naturalness MOS of 3.6/5 and a speaker-similarity SMOS of 4.0/5 — comparable to major English corpora, and openly released to accelerate Persian speech research.

ParsVoice data pipeline: audiobooks are boundary-detected, transcribed, sentence-checked, and boundary-optimized; then quality-filtered, speaker-clustered, punctuation-restored, and assembled into the final ParsVoice corpus. — The nine-stage pipeline that turns raw audiobooks into ParsVoice.

1,804h

High-quality audio

470+

Speakers

3.6/5

Naturalness (MOS)

4.0/5

Speaker similarity

Sample Demonstrations

Zero-shot clones from the fine-tuned model, next to the unseen reference voice each was cloned from.