CSM-1B TTS API (Containerized Developer Edition)

An enhanced, container-ready version of the original CSM-1B TTS API, designed for local development and deployment using VS Code Dev Containers, Docker Desktop, and NVIDIA GPU acceleration.

This fork preserves all features from the original project while adding:

• Developer-friendly containerization via .devcontainer/
• Hot-reload support for code changes
• Safe .env.template handling for Hugging Face tokens
• Portable, reproducible setup for local GPU inference
• Delayed model download (at runtime only, not at build-time)

This setup is intended for developers and researchers who want a local, modifiable version of the CSM-1B TTS API with full GPU support.

• OpenAI-compatible TTS API (drop-in replacement for OpenAI TTS endpoint)
• 6 high-quality voices with consistent acoustic profiles
• Full voice cloning pipeline via Web UI or API (including YouTube input)
• GPU acceleration via PyTorch + CUDA
• Ready-to-run containerized development setup
• Local-first design with persistent voice and audio cache support

This repository is based on a fork of the open-source implementation by community contributor @phildougherty:

• Original Source: https://github.com/phildougherty/sesame_csm_openai
• Maintained Fork: https://github.com/hierro/sesame_csm_openai

Licensing, acknowledgments, and model terms remain unchanged (MIT License).

• Developers working on TTS or AI voice systems
• Researchers exploring voice consistency, cloning, or inference tuning
• Builders of local/private TTS solutions (without cloud dependencies)
• Anyone wanting a reproducible, containerized version of CSM-1B to run locally with NVIDIA GPU acceleration

• Host OS: Windows 10/11 (with WSL2), Linux, or macOS
• Docker:
  • Docker Desktop must be running before you open the project in VS Code
  • WSL2 backend enabled on Windows
• GPU Support:
  • NVIDIA GPU with 8GB+ VRAM (tested with RTX 2080)
  • CUDA 12.8 compatible driver
  • NVIDIA Container Toolkit installed
• VS Code + Extensions:
  • Remote - Containers
  • Docker (optional but helpful)

git clone https://github.com/hierro/sesame_csm_openai.git
cd sesame_csm_openai

cp .env.template .env

Edit .env and add your Hugging Face token:

HF_TOKEN=your_huggingface_token_here

Make sure you’ve accepted model access: https://huggingface.co/sesame/csm-1b

• Open VS Code.
• Run the command:

Dev Containers: Open Folder in Container...

• Select the root folder of the cloned repo.

This will:

• Build the container (without downloading the model yet)
• Install Python and system dependencies
• Mount your repo as /app
• Enable GPU access via Docker Compose

📌 Important: Wait until VS Code finishes installing extensions (e.g. ms-python.python) and you see logs like:

[Run in container: cat /proc/environ]

Then, proceed to the next step.

Open a new terminal inside the container and run:

uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

On first run, the model (6GB+) will be downloaded from Hugging Face. This may take 8–10 minutes.

Once complete, you can open:

• Swagger UI: http://localhost:8000/docs
• Voice Cloning UI: http://localhost:8000/voice-cloning

• Model checkpoint is no longer downloaded during container build. It is now deferred to runtime for efficiency and clarity.
• Audio generation speed and quality depends on GPU and system RAM
• Voice quality improves with usage (voice memory accumulates)
• For better testing, use .wav or .flac inputs
• You can mount volumes or use bind mounts to persist generated content

• Dockerfile and docker-compose.yml (optimized for GPU)
• .devcontainer/ folder for VS Code integration
• .env.template (configurable)
• Persistent folders: /models, /voice_profiles, /voice_memories, etc.
• app/ folder with TTS API and Web UI

• CSM-1B model by Sesame
• Original implementation: @phildougherty
• Hugging Face model: https://huggingface.co/sesame/csm-1b

This fork is not affiliated with Sesame or the original authors. It is intended as a developer-friendly variant for easier local testing, modification, and GPU deployment.

See the Usage Guide for full details on:

• TTS generation (/v1/audio/speech)
• Voice cloning (via API or UI)
• YouTube-based voice cloning
• Conversation API support
• OpenWebUI integration
• Advanced options and troubleshooting