Ghazouaniwala/silma-tts-derja-v2-1

SILMA TTS — Tunisian Derja fine-tune (v2, quality-focused)

Fine-tune of silma-ai/silma-tts on the LinTO Tunisian dataset.

v2 recipe: speaker_mode=multi (single≈AbdelAzizErwi), moderate text normalization, audio-quality filtering (SNR≥8.0dB), phonemization=False, 15 epochs. Targets pronunciation/vowel/consonant quality by improving the signal rather than the architecture. Deploy raw or EMA weights per the A/B in §8.

Usage

Load with the F5-TTS v1.1.7 / SILMA pipeline: model.pt + vocab.txt + config.yaml. The vocab.txt here matches this run (SILMA char vocab, or phonemized vocab if piloted).

Attribution (required — CC BY 4.0)

@misc{linagora2024Linto-tn,
  title={LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect},
  author={Hedi Naouara and Jerome Louradour and Jean-Pierre Lorre},
  year={2025}, eprint={2504.02604}, archivePrefix={arXiv}, primaryClass={cs.CL}
}

Base: SILMA TTS (Apache-2.0). Data: LinTO (CC BY 4.0).

Responsible use

Voice cloning requires documented speaker consent. Do not use for deception or impersonation.