What languages does IndexTTS support?

IndexTTS-2.5 supports Chinese, English, Japanese, Spanish and Arabic. Earlier versions (1.0-2) focus mainly on Chinese and English with Pinyin-based pronunciation control.

Do I need a GPU to run IndexTTS?

Effectively yes. The README covers CUDA setup, GPU checks, FP16/BF16 inference, and DeepSpeed acceleration, all of which assume an NVIDIA GPU; CPU-only use isn't documented.

How much reference audio do I need to clone a voice?

A single reference `.wav` file is enough for zero-shot cloning; no fine-tuning or additional training data is required.

Can I run IndexTTS in production?

Yes, there's a documented recipe for serving IndexTTS-2.5 through vLLM for production deployment, in addition to the WebUI and Python API for local or scripted use.

What's the license for IndexTTS?

The repository metadata does not specify a license, so check the repo directly before using it in a commercial or redistributed product.

How does emotion control work?

You can drive emotion via a separate emotional reference audio clip, an 8-value emotion vector, or automatic extraction from the input text itself, with an `emo_alpha` parameter to tune intensity.

IndexTTS: Zero-shot voice cloning TTS

IndexTTS is a zero-shot text-to-speech system built for developers who need to clone a voice from a single reference audio clip and generate speech in multiple languages. It targets people building voice products, dubbing pipelines, audiobook tools, or research on controllable TTS, not casual users looking for a hosted app. The project ships several model generations (IndexTTS through IndexTTS-2.5), each with weights on HuggingFace and ModelScope, a WebUI, a Python API, and a CLI-style inference script.

The current model, IndexTTS-2.5, adds Chinese, English, Japanese, Spanish and Arabic support, plus fine-grained control over emotion, speaking speed, and pronunciation (Pinyin, CMU phonemes, Japanese Kana), while running faster than IndexTTS-2. Earlier versions (1.0, 1.5, 2) remain usable and documented, so you can pick the tradeoff between stability and feature set that fits your project.

Key features

Zero-shot voice cloning: generate speech in a target voice from a single reference .wav file, no fine-tuning required.
Multilingual synthesis: IndexTTS-2.5 covers Chinese, English, Japanese, Spanish and Arabic, with cross-lingual and timbre-emotion disentanglement.
Emotion control: set emotion via a separate emotional reference audio, an 8-dimension emotion vector, or automatic extraction from the input text (use_emo_text).
Emotion intensity tuning: the emo_alpha parameter scales how strongly the emotion reference affects the output (0.0-1.0).
Speaking speed control: the duration_factor parameter slows down or speeds up generated speech (0.5x-2.0x).
Pronunciation control: override pronunciation inline using Pinyin, CMU phonemes, or Japanese Kana annotations.
Performance options: FP16/BF16 inference, DeepSpeed acceleration, and compiled CUDA kernels to cut VRAM use and speed up generation.
Production serving: a documented vLLM recipe for deploying IndexTTS-2.5 at scale.
WebUI and Python API: a Gradio-based web demo alongside a scriptable Python interface for batch or programmatic use.

Ideal use cases

IndexTTS fits projects that need controllable voice cloning from short reference clips: audiobook narration in a specific voice, dubbing or localization pipelines that need cross-lingual output with a consistent speaker identity, character voices for games or interactive fiction with emotion control, and research on disentangling timbre from emotion in TTS. It's also a reasonable base for a self-hosted TTS backend served via vLLM in production.

It's not a good fit if you want a plug-and-play cloud API with no setup: you need a GPU, model checkpoints (several GB), and a Python environment. It's also not built for real-time, low-latency conversational TTS out of the box, since inference speed depends heavily on hardware and settings like DeepSpeed and half-precision, which can vary in effect. If you need voice cloning with no reference audio consent workflow, be aware this is a raw research/engineering tool without built-in safeguards, so you're responsible for how you source and use reference voices.

Installation

IndexTTS uses uv for dependency management. Clone the repo, install dependencies, download a model, and check GPU availability:

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

Download model weights via huggingface-cli or modelscope:

uv tool install "huggingface-hub"
hf download IndexTeam/IndexTTS-2.5 --local-dir=checkpoints

Check GPU acceleration:

uv run tools/gpu_check.py

Run the web demo:

uv run webui.py

Then open http://127.0.0.1:7860 in a browser. For scripted or production use, call the Python API directly:

from indextts.infer_v2_5 import IndexTTS2
tts = IndexTTS2(cfg_path="checkpoints/config.yaml", model_dir="checkpoints", use_bf16=True)
tts.infer(spk_audio_prompt='examples/voice_01.wav', text="Hello world", lang="EN", output_path="gen.wav")

On Windows, DeepSpeed can be hard to install; skip --all-extras and add other extras manually if needed. A CUDA Toolkit version 12.8 or newer is required if you hit CUDA errors during setup.