- Developer
- bilibili IndexTTS
- Version covered
- v2.0.0 / IndexTTS-2
- Emotion inputs
- Audio, eight-value vector, text
- Python entry point
- indextts.infer_v2.IndexTTS2
Strengths worth exploring
Direct identity and emotion independently
A speaker reference supplies timbre, while an emotional reference supplies style. This allows new combinations without finding a recording of that speaker performing every desired emotion.
Official source [2]Mix eight emotion intensities
The public interface includes happy, angry, sad, afraid, disgusted, melancholic, surprised, and calm dimensions. A vector gives explicit control over the mix.
Official source [1]Give emotion its own text direction
Infer emotion from the script or supply a separate description, so dialogue and acting notes can be edited independently.
Official source [1]Separate research results from released features
The paper proposes precise duration control, but the v2.0.0 release explicitly says this feature is not enabled. A claim that this public release can accept an exact number of seconds for video synchronization would be misleading.
Official source [1]Tune emotion around the line
The released examples expose emo_audio_prompt, emo_vector, and emo_text. Stronger emotion still needs to be checked for clarity and continuity; the guide recommends moderate strengths for text-based emotion modes.
Mixed Chinese/pinyin pronunciation control remains available for valid entries in the model's pinyin vocabulary.
Official source [1]Usage and deployment
- Install Git, Git LFS, and uv, then prepare the supported GPU/PyTorch environment. The commands pin v2.0.0 to keep this guide aligned with its examples.
- Sync WebUI dependencies and download IndexTeam/IndexTTS-2 into checkpoints. Initial startup may also retrieve auxiliary models.
- Start uv run webui.py, choose a speaker reference, and test the desired emotion input. Use the release's IndexTTS2 examples for code integration.
git clone --branch v2.0.0 https://github.com/index-tts/index-tts.git
cd index-tts
git lfs install
git lfs pull
uv sync --extra webui
uv tool install 'huggingface-hub[cli]'
hf download IndexTeam/IndexTTS-2 --local-dir checkpoints
uv run webui.pyBefore you get started
- Do not treat v2.0.0, later 2.x releases, and every paper experiment as the same runnable feature set.
- The examples warn that random sampling may reduce voice similarity. Compare identity and speech clarity when changing emotion controls.
Where it fits
Useful for game characters, emotional dialogue, audiobooks, and workflows where one consistent speaker performs several moods.
Official resources and sources
DS Speech currently provides Qwen-Audio-3.0-TTS online. Other models in this library are documented for learning and self-deployment.