yeowon/index-tts-ko

IndexTTS-KO: 180,100-Step Korean Zero-Shot Voice & Emotion Cloning TTS

🇺🇸 English  |  🇰🇷 한국어  |  🇨🇳 中文  |  🇯🇵 日本語

🏦 국내 계좌이체 후원 (Korean Bank Transfer): NH농협(농축협) 352-1556-0640-33 (예금주: 윤여원)  |  💛 카카오페이 QR 송금  |  ☕ Buy Me a Coffee

🎬 Audio Comparison Demo: Reference Voice vs Base IndexTTS-2 (Untrained Korean) vs IndexTTS-KO (180,100 Steps, 16k Korean BPE + Emotion Cloning — Turn Sound ON 🔊)

🔓 100% Ungated Open Release (Zero-Gate Access): All checkpoint weights (gpt.pth, s2mel.pth, qwen0.6bemo4-merge, 16,000 Korean BPE tokenizer) are openly downloadable without manual approval or login tokens. ⚡ Want another language finetuned or a large model quantized (Rung Ω Base-243 + TheoremFlow-Ω)? Leave a request on our Community Board (Discussions #1) or get Priority Queue processing via Buy Me a Coffee (buymeacoffee.com/tripleyoung)!

yeowon/index-tts-ko is a Korean-finetuned release of IndexTeam/IndexTTS-2 trained for 180,100 steps on emilia-yodas/KO_preprocessed. Given a short reference audio prompt (reference.wav) and Korean text, it synthesizes natural 22.05 kHz Korean speech while cloning speaker timbre and emotional prosody. Inference runs with edu-ide/index-tts.

This release extracts the clean inference-only weights from best_model.pth at step 180,100 (optimizer states stripped, torch.compile _orig_mod. prefixes removed, original dtypes preserved, and all 665 GPT tensors verified for bit-exact equality against the trained checkpoint).

---

🚀 Quickstart (Linux & Python 3.10)

Install Git and uv, then run:

git clone https://github.com/edu-ide/index-tts.git
cd index-tts
git checkout 80060cced8cd0015dfaaaadfc70509fdec5bfb1f
uv python install 3.10
uv sync --python 3.10 --frozen --no-dev
export HF_MODEL_ID="yeowon/index-tts-ko"

Place a reference voice clip at reference.wav in the current directory and run FP16 synthesis:

.venv/bin/python - <<'PY'
import os
from pathlib import Path

from huggingface_hub import snapshot_download
from indextts.infer_v2 import IndexTTS2

model_dir = Path(snapshot_download(repo_id=os.environ["HF_MODEL_ID"]))

tts = IndexTTS2(
    cfg_path=str(model_dir / "config.yaml"),
    model_dir=str(model_dir),
    use_fp16=True,
    use_cuda_kernel=False,
    use_deepspeed=False,
)
tts.normalizer.preferred_language = "ko"
tts.infer(
    spk_audio_prompt="reference.wav",
    text="안녕하세요. 한국어 음성 합성 모델입니다. 오늘 오후 세 시에 만나요.",
    output_path="output.wav",
)
PY

Key Usage Notes

• tts.normalizer.preferred_language = "ko": Ensures Korean text is routed through the Korean text normalizer rather than CJK/English fallback paths (infer(language="ko") is not a separate argument). Spell out numbers and units in Korean Hangul as you want them pronounced.
• Dedicated 16,000 Korean BPE Tokenizer: Always use the bundled config.yaml and bpe.model together. Mixing configs or tokenizers from the base IndexTTS-2 checkpoint will cause vocabulary dimension mismatches.
• Automatic Auxiliary Downloads on First Run:
  • facebook/w2v-bert-2.0
  • amphion/MaskGCT
  • funasr/campplus
  • nvidia/bigvgan_v2_22khz_80band_256x

---

📦 Included Files & SHA-256 Verified Artifacts

File	Size / Role
gpt.pth (3.53 GB)	Step-180,100 Korean-finetuned autoregressive GPT weights (665 tensors verified, optimizer stripped)
config.yaml, bpe.model	Model architecture configuration & 16,000-token Korean BPE tokenizer
s2mel.pth (1.20 GB)	Semantic-to-Mel spectrogram synthesis weights
wav2vec2bert_stats.pt	Speech feature normalization statistics
feat1.pt, feat2.pt	Speaker timbre & emotion conditioning feature matrices
qwen0.6bemo4-merge/ (1.19 GB)	Merged Qwen-0.6B emotion conditioning encoder & tokenizer
release.json, smoke-test.json	Tensor equality verification manifest (SHA-256) & RTX 4090 FP16 smoke-test report
LICENSE, LICENSE_ZH.txt	Original bilibili Model Use License Agreement

---

🔬 Training Data & Verification Scope

• Training Manifest: emilia-yodas/KO_preprocessed derived from Emilia / Emilia-YODAS and YODAS. Raw training audio is not redistributed in this repository.
• Verification: Tensor equality before/after checkpoint extraction (665 GPT keys) and tokenizer/config alignment were verified. On an NVIDIA GeForce RTX 4090 (FP16, use_cuda_kernel=False, use_deepspeed=False), synthesizing "안녕하세요. 한국어 음성 합성 테스트입니다." with examples/voice_01.wav produced a finite, non-silent 22,050 Hz (3.61s, peak 0.632) WAV waveform (smoke-test.json). Formal MOS/CER human evaluations are not claimed in smoke-test.json.

📜 License & Derivative Work Notice

This model is governed by the original IndexTTS-2 bilibili Model Use License Agreement (English LICENSE / Chinese LICENSE_ZH.txt).

Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work.

该衍生品对原模型所作的任何改动与原模型原始权利人无关，原始权利人对该衍生品不背书、不担保、不承担责任。

---

☕ Support Our Open-Source Research & Priority Model Queue

We release IndexTTS-KO and our Rung Ω / TheoremFlow-Ω models with zero gated barriers. If this model helps your Korean voice workflow or research:

1. ⚡ Priority Supporter Queue: Leave a target model or language request in your Buy Me a Coffee support note (buymeacoffee.com/tripleyoung) or KakaoPay/Bank transfer memo — supporter requests are prioritized first!
2. 💬 Community Request Board: Open a thread anytime in Hugging Face Discussions.

• ☕ Buy Me a Coffee (Global Support & Priority Requests): https://buymeacoffee.com/tripleyoung
• 💛 KakaoPay QR Transfer (한국 카카오페이 QR 송금 후원): https://qr.kakaopay.com/FVgCIrWbX
• 🏦 Korean Bank Transfer (국내 계좌이체 후원): NH농협(농축협) 352-1556-0640-33 (예금주: 윤여원)
• 💬 Community & Model Request Board: https://huggingface.co/yeowon/index-tts-ko/discussions