miso-tts-docker

miso-tts-docker is an unofficial Docker Compose package designed to facilitate running the Miso TTS 8B speech synthesis model locally on Windows systems, particularly those equipped with NVIDIA RTX 50-series (Blackwell) GPUs. It addresses common challenges faced when deploying research models on Windows with Docker Desktop and new GPU architectures, such as compatibility issues with tokenizer access, audio I/O dependencies, and launch script portability.

The package provides GPU-ready containers that wrap the upstream Miso TTS 8B inference, enabling high-quality local speech generation for use cases like agents and demonstrations. It includes persistent Hugging Face cache volumes, preflight checks for gated Llama 3.2 tokenizer access, and a patch for soundfile audio I/O to avoid reliance on torchcodec with PyTorch 2.11. The default configuration is optimized for full bfloat16 support and recommends approximately 24 GB of VRAM for optimal performance. Windows users benefit from provided .cmd launchers, which are intended to simplify startup and avoid PowerShell execution policy issues. The Docker Compose setup incorporates NVIDIA GPU passthrough and is oriented for use with PyTorch 2.11 and CUDA 12.8, specifically targeting the RTX 5090 and related Blackwell architecture cards.

This solution is intended for developers and builders who need a streamlined, clone-and-run approach to deploying the Miso TTS 8B model on their own Windows hardware, particularly when existing research repositories are difficult to adapt to the Windows and Docker Desktop environment. The package is not affiliated with Miso Labs, and its primary contribution is making the upstream model accessible in a Windows-centric workflow with modern NVIDIA GPUs.

evidence_sufficient": true}