ParsVoice XTTS-v2 (Persian)

A Persian (Farsi) text-to-speech model fine-tuned from Coqui XTTS-v2 on the ParsVoice corpus, a large-scale multi-speaker Persian speech dataset. The model supports zero-shot voice cloning: given a short reference clip of a target speaker, it synthesizes Persian speech in that speaker's voice.

This model accompanies the paper ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis (EMNLP 2026, Main Conference).

Model Details

Base model: Coqui XTTS-v2
Language: Persian (fa)
Task: Zero-shot, multi-speaker text-to-speech / voice cloning
Training data: ParsVoice (2,200 hours of TTS-ready audio, 1,800+ speakers)
Fine-tuned by: Mohammad Javad Ranjbar Kalahroodi

Evaluation

Evaluated on held-out Persian speakers:

Naturalness (MOS): 3.6 / 5
Speaker similarity (SMOS): 4.0 / 5

See the paper for the full evaluation protocol and comparisons.

Usage

from TTS.api import TTS

tts = TTS("MohammadJRanjbar/parsvoice-xtts-v2")

tts.tts_to_file(
    text="سلام، حال شما چطور است؟",
    speaker_wav="reference_speaker.wav",
    language="fa",
    file_path="output.wav",
)

Provide a clean reference clip (roughly 6 to 15 seconds, 16 kHz or higher, minimal background noise) for best speaker cloning results.

Training

Fine-tuning setup: [e.g. full fine-tune / adapter, number of epochs, batch size, hardware]
Preprocessing: Audio filtered and quality-scored through the ParsVoice pipeline (ASR transcription, sentence-completion filtering, boundary optimization, ECAPA-TDNN speaker clustering, multi-dimensional quality scoring).

Limitations

Performance depends on the quality of the reference clip; noisy or very short references degrade output.
The model reflects the speaker and domain distribution of ParsVoice and may generalize less well to out-of-distribution accents, dialects, or speaking styles.
Numbers, code-switched text, and rare named entities may be mispronounced.
As a generative voice-cloning model, it can be misused to imitate real people's voices. Please use responsibly and with consent from any speaker being cloned.

License

This model is released under the Coqui Public Model License (CPML), inherited from the XTTS-v2 base model. The CPML restricts commercial use; review the full license before use.

Citation

If you use this model or the ParsVoice corpus, please cite:

@inproceedings{ranjbar2026parsvoice,
  title     = {ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis},
  author    = {Ranjbar Kalahroodi, Mohammad Javad and <FILL IN CO-AUTHORS>},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  address   = {Budapest, Hungary}
}

Acknowledgements

This work began as a Master's thesis at the University of Tehran, supervised by Dr. Azadeh Shakery and Dr. Heshaam Faili.

MohammadJRanjbar/ParsVoice-XTTS