xchaix74/chatterbox-tts-api-blackwell repository overview

⁠Chatterbox TTS API — Blackwell Edition

Chatterbox Multilingual TTS served over an OpenAI-compatible API. Built specifically for NVIDIA RTX 50-series (Blackwell, sm_120) GPUs.

⁠Why this image exists

The upstream travisvn/chatterbox-tts-api:gpu image ships PyTorch 2.6.0 + CUDA 12.4, which does not have compiled kernels for Blackwell's sm_120 compute capability. Loading the model fails with:

RuntimeError: CUDA error: no kernel image is available for execution on the device
RuntimeError: operator torchvision::nms does not exist

This image rebuilds the stack on PyTorch 2.7.1 + CUDA 12.8, which has native sm_120 support. It's the fix I needed, published so others don't have to solve it again.

⁠Quick start

docker run --rm --gpus all \
  -p 8307:4123 \
  -e HF_TOKEN=your_huggingface_token \
  -v chatterbox-cache:/cache \
  xchaix74/chatterbox-tts-api-blackwell:latest

Synthesize:

curl -X POST http://localhost:8307/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello from Chatterbox", "voice": "default"}' \
  -o out.wav

Health check:

curl http://localhost:8307/health

⁠Environment variables

• HF_TOKEN — required. Get one at https://huggingface.co/settings/tokens⁠ (The Chatterbox model is gated; the token is needed to download weights.)
• DEVICE — cuda (default) or cpu

⁠Volume

• /cache — Hugging Face model cache. Mount a named volume so weights persist across container restarts (~2 GB download on first run).

⁠Tested on

• NVIDIA RTX 5060 Ti (sm_120), 16 GB VRAM
• Driver 570.x, CUDA 12.8
• Ubuntu 22.04 host

⁠Supported languages

22 languages including: English, Malay, Arabic, Mandarin, Japanese, Korean, Hindi, Tamil, Spanish, French, German, Italian, Portuguese, Russian, and more.

⁠License

• Chatterbox model: MIT License (Resemble AI)
• API wrapper: MIT (travisvn)
• Runtime includes NVIDIA CUDA libraries, subject to NVIDIA's CUDA EULA. This software contains source code provided by NVIDIA Corporation.

⁠Credits

• Chatterbox by Resemble AI: https://github.com/resemble-ai/chatterbox⁠
• API wrapper by travisvn: https://github.com/travisvn/chatterbox-tts-api⁠
• Blackwell build and packaging by xchaix74