index-tts/index-tts

📚 What is IndexTTS?

IndexTTS is an open‑source, zero‑shot text‑to‑speech (TTS) system that can clone a voice from just a single reference audio clip. The latest version, IndexTTS‑2.5, works in five languages (Chinese, English, Japanese, Spanish, Arabic) and lets you control:

Emotion – via an emotion‑reference audio, an explicit emotion vector, or automatically from the text.
Speaking speed – with a duration_factor (0.5×‑2.0×).
Pronunciation – by supplying Pinyin, CMU phonemes, or Japanese Kana directly in the input text.
Voice timbre – the cloned voice retains the speaker’s timbre while you can still change language.

The model is positioned as an “industrial‑level” solution: it runs fast enough for real‑time use, supports half‑precision (BF16/FP16) inference, and can be served at scale with the vLLM recipe.

🚀 Getting Started (Quick‑Start)

Clone the repo

git clone https://github.com/index-tts/index-tts.git && cd index-tts

Install the environment – the project uses the modern Python package manager uv.
```
pip install -U uv          # install uv itself
uv sync --all-extras       # creates a .venv and installs all deps
```
Optional: add a Chinese mirror (--default-index) if you have a slow connection.
Download a model (e.g., the newest IndexTTS‑2.5) via the Hugging Face CLI:
```
uv tool install huggingface-hub
hf download IndexTeam/IndexTTS-2.5 --local-dir=checkpoints
```
(You can also use modelscope-hub with the same command pattern.)
Run the web UI (the simplest way to test the system):
```
uv run webui.py          # defaults to IndexTTS‑2.5
# or, for the older 2.0 model:
uv run webui.py --version 2 --model_dir ./checkpoints_2
```
Open http://127.0.0.1:7860 in a browser. The UI lets you upload a reference voice, type text, pick a language, and tweak the controls described above.

🛠️ Core Features

Feature	How it works	Example flag / API call
Zero‑shot voice cloning	A single reference wav is fed to the model as `spk_audio_prompt`.	`tts.infer(spk_audio_prompt='voice.wav', text='Hello')`
Emotion control	Provide an emotion‑reference wav (`emo_audio_prompt`), an 8‑dim emotion vector (`emo_vector`), or let the model infer emotions from the text (`use_emo_text`).	`emo_alpha=0.6`, `emo_vector=[0,0,0.8,0,0,0,0,0]`
Speed control	Scale the synthesis duration with `duration_factor` (0.5‑2.0).	`duration_factor=1.2` (slow)
Pronunciation control	Insert `<phoneme	X>` tags in the input text. Supports Chinese Pinyin, English CMU phonemes, Japanese Kana.
Half‑precision inference	BF16 for IndexTTS‑2.5, FP16 for IndexTTS‑2 – reduces VRAM and speeds up generation.	`use_bf16=True` (constructor)
DeepSpeed / compiled CUDA kernels	Optional acceleration for compatible GPUs.	`--extra deepspeed` during install
Production serving	A ready‑made vLLM recipe (`recipes.vllm.ai/IndexTeam/IndexTTS-2.5`).	See the linked recipe page

📦 Model Zoo (pre‑trained checkpoints)

Model	Languages	Demo	Model download (HF / ModelScope)
IndexTTS‑2.5	ZH, EN, JA, ES, AR	[Demo page]	`IndexTeam/IndexTTS-2.5`
IndexTTS‑2	ZH, EN	[Demo page]	`IndexTeam/IndexTTS-2`
IndexTTS‑1.5	ZH, EN	[Demo page]	`IndexTeam/IndexTTS-1.5`
IndexTTS‑1	ZH, EN	[Demo page]	`IndexTeam/Index-TTS`

🧑‍💻 Using the Python API (code snippet)

# Initialise (choose the version you downloaded)
from indextts.infer_v2_5 import IndexTTS2   # for 2.5
tts = IndexTTS2(
    cfg_path='checkpoints/config.yaml',
    model_dir='checkpoints',
    use_bf16=True,          # half‑precision
)

# Simple voice clone
tts.infer(
    spk_audio_prompt='examples/voice_01.wav',
    text='Hello world',
    lang='EN',
    output_path='gen.wav',
    verbose=True,
)

# Emotion + speed control
tts.infer(
    spk_audio_prompt='examples/voice_07.wav',
    text='这是一段情绪化的中文',
    lang='ZH',
    emo_audio_prompt='examples/emo_sad.wav',
    emo_alpha=0.8,
    duration_factor=0.9,
    output_path='emo.wav',
)

The same pattern works for the older IndexTTS2 class (indextts.infer_v2).

📊 Reported Performance

The authors provide two benchmark tables on the CV3‑Eval suite (multilingual speech quality and cross‑lingual synthesis). Highlights for the flagship IndexTTS‑2.5‑RL (reinforced‑learning variant):

Average word‑error‑rate (WER) ≈ 6 % across five languages.
Average subjective score (SS) ≈ 73 / 100, comparable to larger models (e.g., VoxCPM2 with 2 B parameters).
Model size: 0.8 B parameters – considerably smaller than many competing systems.

🛡️ License & Community

License: The repository is released under the Apache‑2.0 license (as indicated in the repo’s LICENSE file). This permits commercial use, modification, and redistribution.
Community: A Discord server is linked for support, and the project maintains a Discord badge in the README. Issues and pull requests are accepted via GitHub.

📚 When to Use IndexTTS?

Rapid prototyping of voice‑cloned speech in multiple languages.
Interactive applications (e.g., virtual assistants, game characters) where you need low‑latency TTS with controllable emotion and speed.
Research on zero‑shot multilingual TTS, voice cloning, or controllable speech synthesis.
Production – the vLLM recipe and half‑precision inference make it feasible to serve at scale.

🎯 TL;DR

IndexTTS is a real, open‑source, zero‑shot multilingual TTS engine that clones a voice from a single audio clip and offers fine‑grained control over emotion, speed, and pronunciation. It ships with ready‑to‑use pretrained models, a simple uv‑based install, a web UI, and a Python API, and it can be deployed in production via vLLM. The project is actively maintained (latest release 2.5 in Aug 2026) and licensed under Apache‑2.0.