indextts2-inference 2.1.0

IndexTTS2 inference library: zero-shot text-to-speech with emotional control

index-tts-inference

Minimal pip package for IndexTTS2 inference. Wraps the official IndexTTS2 repo, stripped down to only what's needed for inference.

Install

pip install indextts2-inference

Optional extras

SageAttention (alternative attention backend):

pip install indextts2-inference[sage-attn]

Flash Attention v2 (acceleration engine with KV cache and CUDA graphs):

pip install indextts2-inference[flash-attn]

DeepSpeed:

pip install indextts2-inference[deepspeed]

Usage

from indextts import IndexTTS2

# Auto-downloads model from HuggingFace
tts = IndexTTS2()

# Or use local/finetuned checkpoints
tts = IndexTTS2(model_dir="/path/to/checkpoints")

# Basic inference
tts.infer(spk_audio_prompt="voice.wav", text="Hello world", output_path="out.wav")

Attention backends

# Default PyTorch SDPA — auto-selects best kernel, no extra deps needed
tts = IndexTTS2()

# SageAttention — may help on Ampere/Hopper GPUs, requires sageattention package
tts = IndexTTS2(attn_backend="sage", use_fp16=True)

# Flash Attention v2 — acceleration engine with paged KV cache and CUDA graphs
tts = IndexTTS2(attn_backend="flash")

Language selection

By default the language is auto-detected between Chinese and English. You can set it explicitly:

tts = IndexTTS2(language="es")
tts.infer(spk_audio_prompt="voice.wav", text="Hola, esto es una prueba.", output_path="out.wav")

Emotion control

There are three ways to control the emotion of the generated speech:

# 1. From a reference audio
tts.infer(
    spk_audio_prompt="speaker.wav",
    text="Some text",
    output_path="out.wav",
    emo_audio_prompt="happy_reference.wav",
    emo_alpha=0.7,
)

# 2. With an explicit emotion vector
#    [happy, angry, sad, afraid, disgusted, melancholic, surprised, calm]
tts.infer(
    spk_audio_prompt="speaker.wav",
    text="I am very happy!",
    output_path="out.wav",
    emo_vector=[0.8, 0, 0, 0, 0, 0, 0, 0],
)

# 3. Auto-detect emotion from the text itself
tts.infer(
    spk_audio_prompt="speaker.wav",
    text="I am very happy!",
    output_path="out.wav",
    use_emo_text=True,
)

Streaming

for chunk in tts.infer(
    spk_audio_prompt="voice.wav",
    text="Long text to synthesize...",
    output_path="out.wav",
    stream_return=True,
):
    if chunk is not None and hasattr(chunk, "shape"):
        audio_np = chunk.squeeze().cpu().numpy()

Generation parameters

You can tune sampling parameters via kwargs:

tts.infer(
    spk_audio_prompt="voice.wav",
    text="Hello",
    output_path="out.wav",
    temperature=0.6,
    top_k=20,
    top_p=0.8,
    max_mel_tokens=2000,
)

Logging

By default, index-tts-inference only shows warnings. To see detailed logs:

export INDEXTTS_LOG_LEVEL=DEBUG  # DEBUG, INFO, WARNING (default)

PyTorch with CUDA

This package lists torch and torchaudio as dependencies without pinning a specific CUDA version. Install the CUDA variant you need before installing this package:

# Example: PyTorch with CUDA 12.8
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128

# Then install the package
pip install indextts2-inference

Or with uv:

# pyproject.toml of your project
[tool.uv.sources]
torch = [{ index = "pytorch-cuda", marker = "sys_platform == 'linux'" }]
torchaudio = [{ index = "pytorch-cuda", marker = "sys_platform == 'linux'" }]

[[tool.uv.index]]
name = "pytorch-cuda"
url = "https://download.pytorch.org/whl/cu128"
explicit = true

License

See LICENSE and DISCLAIMER.