🎙️ Qwen3-TTS

A powerful Text-to-Speech application featuring voice cloning, voice design, and custom voice generation powered by Alibaba's Qwen3-TTS models.

✨ Features

• 🎨 Voice Design - Create custom voices using natural language descriptions
• 🎭 Voice Clone - Clone any voice from a reference audio sample
• 🗣️ Custom Voice - Generate speech with predefined speakers and style instructions
• 📝 Long Text Chunking - Automatically splits long text at sentence/word boundaries
• 🎤 Whisper Transcription - Auto-transcribe reference audio for voice cloning
• ⚙️ Model Management - Download, load, and unload models with a simple UI

🖥️ Requirements

• Python 3.10+
• NVIDIA GPU with CUDA 12.8 support
• ~8GB+ VRAM for 0.6B models, ~16GB+ for 1.7B models

📦 Installation (Windows)

1. Clone the repository

git clone https://github.com/SUP3RMASS1VE/Qwen3-TTS.git
cd Qwen3-TTS

2. Create and activate virtual environment

python -m venv venv
venv\Scripts\activate

3. Install uv (fast package installer)

pip install uv

4. Install dependencies

uv pip install -r requirements.txt

5. Install PyTorch with CUDA 12.8

uv pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128

6. Install Triton for Windows

uv pip install triton-windows==3.3.1.post19

7. Install Flash Attention (optional but recommended)

uv pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.10/flash_attn-2.8.2+cu128torch2.7-cp310-cp310-win_amd64.whl

🚀 Usage

Start the application

python app.py

The Gradio UI will launch at http://localhost:7860

Available Models

Model Type	Sizes	Description
VoiceDesign	1.7B	Create voices from text descriptions
Base	0.6B, 1.7B	Voice cloning from reference audio
CustomVoice	0.6B, 1.7B	TTS with predefined speakers

Tips

• Models auto-download on first use, or pre-download them in the Models tab
• Use the Chunk Size slider for long texts (splits at sentence boundaries)
• Click 🔄 Refresh Status to see models loaded from other tabs
• Whisper auto-unloads after transcription to free GPU memory

📁 Project Structure

Qwen3-TTS/
├── app.py                 # Main Gradio application
├── requirements.txt       # Python dependencies
├── README.md             # This file
└── qwen_tts/             # Core TTS module
    ├── core/             # Model configurations and implementations
    │   ├── models/       # Qwen3-TTS model code
    │   ├── tokenizer_12hz/
    │   └── tokenizer_25hz/
    └── inference/        # High-level inference wrappers

🎛️ Supported Languages

Auto, Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian

🎭 Available Speakers (CustomVoice)

Aiden, Dylan, Eric, Ono_anna, Ryan, Serena, Sohee, Uncle_fu, Vivian

🙏 Credits

• Qwen3-TTS by Alibaba Qwen Team
• OpenAI Whisper for transcription
• Gradio for the web UI

📄 License

This project uses models from Alibaba's Qwen team. Please refer to the original Qwen3-TTS repository for model licensing information.