prince4332/dots.tts-mf-akan-twi

dots.tts-mf fine-tuned on Akan (Twi)

Fine-tuned from `dots-studio/dots.tts-mf` (the MeanFlow-distilled, low-latency checkpoint of `studio-dots-ai/dots.tts`) on `ghanaopenai/asante-twi-bible-speech-text` (~34.2k Asante Twi speech/text pairs).

• —Fine-tuned with the repo's own scripts/train_dots_tts.py, starting from configs/dots_tts.yaml.
• —Only train.pretrained_model_path, train_data.sources/val_data.sources, train.output_dir, and train.max_train_steps were overridden — batch size, learning rate, optimizer, scheduler, and all other hyperparameters are the repo's shipped defaults.
• —Recommended --num-steps 4 at inference, matching the MeanFlow-distilled base checkpoint.

License: the training data is cc-by-nc-4.0 (non-commercial). This checkpoint inherits that restriction — research / non-commercial use only. dots.tts code/checkpoints themselves are Apache-2.0; see the dots.tts risks & limitations notes on voice-cloning misuse before deploying.

Reference: dots.tts Technical Report, arXiv:2606.07080.