ngoctham/IndexTTS-2.5

IndexTTS-2.5

IndexTTS-2.5 is a zero-shot text-to-speech model that performs voice cloning from a single reference audio. It supports Chinese, English, Japanese, Spanish, and Arabic, with cross-lingual voice transfer and disentangled emotion control.

Compared to IndexTTS-2, it adds Japanese/Spanish/Arabic support, improves inference speed, and enhances controllability of Chinese Pinyin, English CMU phonemes, and Japanese Kana.

Model Details

• —Developed by: IndexTeam, Bilibili
• —Model type: Autoregressive zero-shot TTS (GPT + DiT + BigVGAN)
• —Languages: Chinese, English, Japanese, Spanish, Arabic
• —License: Bilibili Model License

Model Sources

• —Repository: github.com/index-tts/index-tts
• —Demo: IndexTTS-2.5 Demo Page

How to Get Started

Installation

git clone https://github.com/index-tts/index-tts.git && cd index-tts
git lfs pull
pip install -U uv
uv sync --all-extras

Download Model Weights

# HuggingFace
uv tool install "huggingface-hub[cli,hf_xet]"
hf download IndexTeam/IndexTTS-2.5 --local-dir=checkpoints

# ModelScope
uv tool install "modelscope"
modelscope download --model IndexTeam/IndexTTS-2.5 --local_dir checkpoints

Quick Inference

from indextts.infer_v2_5 import IndexTTS2

tts = IndexTTS2(cfg_path="checkpoints/config.yaml", model_dir="checkpoints", use_bf16=True)

# Basic voice cloning
tts.infer(
    spk_audio_prompt="prompt.wav",
    text="Hello, this is a voice cloning demo.",
    lang="EN",
    output_path="output.wav",
)

# With emotion control
tts.infer(
    spk_audio_prompt="prompt.wav",
    text="快躲起来！是他要来了！",
    lang="ZH",
    output_path="output.wav",
    emo_vector=[0, 0, 0.8, 0, 0, 0, 0, 0],
)

# With Pinyin/phoneme annotation
tts.infer(
    spk_audio_prompt="prompt.wav",
    text="他在银<行|XING2>里<行|HANG2>走了半天。",
    lang="ZH",
    output_path="output.wav",
)

Web Demo

uv run webui.py --version 2.5 --model_dir ./checkpoints

Uses

Direct Use

• —Zero-shot voice cloning from a single reference audio
• —Multilingual speech synthesis (Chinese, English, Japanese, Spanish, Arabic)
• —Cross-lingual voice transfer (e.g., Chinese speaker voice → English output)
• —Emotion-controllable speech synthesis via emotion vectors, emotion reference audio, or text-based emotion detection

Downstream Use

• —Audiobook and podcast production
• —Voice dubbing and localization
• —Conversational AI and virtual assistants

Out-of-Scope Use

• —Impersonation or deception without consent
• —Generating misleading or fraudulent audio content
• —Any use that violates applicable laws or regulations

Bias, Risks, and Limitations

• —Voice cloning quality may vary across speakers and languages.
• —The model may produce artifacts or unnatural prosody for very long or highly complex text.
• —Cross-lingual transfer quality depends on the target language and speaker characteristics.
• —The model does not verify speaker identity or consent. Users are responsible for ethical use.

Evaluation

Zero-Shot TTS (CV3-Eval)

Cross-Lingual TTS (Chinese prompt → target language)