Talking Snake

PDF and web page to speech using Qwen3-TTS. Upload a document or URL, get it read aloud with 9 natural voices across English, Chinese, Japanese, and Korean. Audio streams progressively while generation continues.

Deploy Your Own

Click the button above to deploy your own GPU-powered instance. You'll be prompted to create a Hugging Face account and select hardware (L4 or A100 recommended for speed, ~$0.80-$4/hr).

Run Locally

Requires Python 3.11+, NVIDIA GPU (~6GB VRAM), and SoX (apt install sox libsox-dev). The GPU will be automatically freed if the app is idle for 5+ minutes. It can also run on CPU (no GPU, but much slower).

uv sync && uv run --no-sync talking-snake --port 8888  # Open http://localhost:8888

Flash Attention (Optional, ~2x faster)

Flash Attention requires matching your CUDA driver version. Check yours with nvidia-smi (top right shows "CUDA Version").

1. Find a prebuilt wheel at flashattn.dev matching your:

   • CUDA version (e.g., cu130 for CUDA 13.0)
   • PyTorch version (e.g., torch2.10)
   • Python version (e.g., cp312 for Python 3.12)

2. Install matching torch, torchaudio, and flash-attn:

   # Example for CUDA 13.0 + PyTorch 2.10 + Python 3.12
   uv pip install torch==2.10.0+cu130 torchaudio==2.10.0+cu130 --index-url https://download.pytorch.org/whl/cu130
   uv pip install <flash-attn-wheel-url>

3. Run with --no-sync to prevent uv from removing the manually installed packages:

   uv run --no-sync talking-snake --port 8888

▶️ Listen to a sample

The website looks like this:

License

This project is licensed under the MIT License. Dependencies and third-party components (e.g., Qwen3-TTS, SoX) are subject to their own licenses.