[Open Release] IndexTTS-KO: 180,100-Step Korean Finetune of IndexTTS-2 (16k Korean BPE, Zero-Shot Voice Cloning, Ungated)
Hi IndexTTS-2 community! π
We have open-sourced yeowon/index-tts-ko, a 180,100-step Korean finetune of IndexTeam/IndexTTS-2 trained on emilia-yodas/KO_preprocessed with a dedicated 16,000-token Korean BPE vocabulary (bpe.model + config.yaml) and clean FP16 inference weights (gpt.pth with all 665 tensors verified).
β¨ Key Highlights
- π 100% Ungated Open Access: Manual approval gate has been completely removed β download and run immediately with
snapshot_download("yeowon/index-tts-ko"). - π°π· Dedicated 16k Korean BPE + Normalizer: Native Korean Hangul synthesis with zero-shot speaker voice cloning & emotion control (
tts.normalizer.preferred_language = "ko"). - π€ Hugging Face Model Card (EN / KO / ZH / JA): https://huggingface.co/yeowon/index-tts-ko
- π» Inference Repository: https://github.com/edu-ide/index-tts
- π¬ Community & Custom Language/Quantization Request Board: https://huggingface.co/yeowon/index-tts-ko/discussions/1
- β Support & Priority Model Requests: Buy Me a Coffee | KakaoPay QR |
NHλν 352-1556-0640-33 (μ€μ¬μ)
I work on CloudGPU. Thanks for sharing the vocabulary and config together with the weights; with a new BPE that is the part people most often get wrong. For anyone comparing this finetune with the base model, it would help to publish a small fixed evaluation set: a handful of reference prompts and speaker clips, the exact checkpoint step, and the normalizer setting used, with the generated audio files. Then others can tell whether a difference they hear comes from the finetune or from a mismatched tokenizer or config on their side.
