mio-tts-0.6b-bg-finetuned

MioTTS-0.6B Bulgarian Fine-Tuned (BG/EN)

Български

Това е fine-tuned версия на Aratako/MioTTS-0.6B, адаптирана за български TTS.

Какво е модифицирано

• Fine-tune на LLM частта за български текст → speech tokens.
• Обучение върху български dataset (24kHz, двама говорители), с финален run върху филтриран subset.
• Запазен е оригиналният tokenizer/архитектура (Qwen3ForCausalLM, vocab 164480).
• Няма промени в архитектурата на модела, само в learned weights.
• За inference е използван MioCodec pipeline (съвместим с Aratako/MioCodec-25Hz-24kHz).

Финален checkpoint

• Избран финален checkpoint: checkpoint-2000 (Run 4).
• Наблюдаван eval loss в края: ~5.176.

Лиценз и оригинален проект

• Лицензът е запазен: Apache-2.0.
• Оригинален проект: Aratako/MioTTS-0.6B.

Важно

• Този модел е предназначен за research/engineering use.
• Моля спазвайте местното законодателство при voice cloning.

---

Samples

<audio controls src="https://huggingface.co/beleata74/mio-tts-0.6b-bg-finetuned/resolve/main/samples/audio.wav"></audio> <audio controls src="https://huggingface.co/beleata74/mio-tts-0.6b-bg-finetuned/resolve/main/samples/audio%20(1).wav"></audio> <audio controls src="https://huggingface.co/beleata74/mio-tts-0.6b-bg-finetuned/resolve/main/samples/audio%20(2).wav"></audio> <audio controls src="https://huggingface.co/beleata74/mio-tts-0.6b-bg-finetuned/resolve/main/samples/audio%20(3).wav"></audio> <audio controls src="https://huggingface.co/beleata74/mio-tts-0.6b-bg-finetuned/resolve/main/samples/audio%20(4).wav"></audio>

---

English

This is a fine-tuned version of Aratako/MioTTS-0.6B, adapted for Bulgarian TTS.

What was modified

• Fine-tuned the LLM component for Bulgarian text → speech-token generation.
• Trained on a Bulgarian dataset (24kHz, two speakers), with the final run on a filtered subset.
• Preserved original tokenizer/architecture (Qwen3ForCausalLM, vocab 164480).
• No architectural changes; only model weights were updated.
• Inference is used with MioCodec pipeline (compatible with Aratako/MioCodec-25Hz-24kHz).

Final checkpoint

• Final selected checkpoint: checkpoint-2000 (Run 4).
• Final observed eval loss: ~5.176.

License and original project reference

• License preserved: Apache-2.0.
• Original project: Aratako/MioTTS-0.6B.

Notes

• This model is intended for research/engineering usage.
• Please follow local laws and policies for voice cloning usage.