High-quality local speech is finally useful for agents and demos — if you can get the GPU stack to boot. miso-tts-docker is our unofficial Docker Compose pack for running Miso TTS 8B on NVIDIA GPUs, with a special focus on Windows, Docker Desktop, and RTX 50-series (Blackwell) cards.
Not affiliated with Miso Labs — we wrap upstream inference with GPU-ready containers, Hugging Face caching, and Windows-friendly launchers.
Why we packaged it
Raw research repos rarely survive first contact with Docker Desktop on Windows or brand-new GPU arch flags. Builders hit tokenizer gatekeeping, torchcodec gaps, cache churn, and “works on my Linux box” launch scripts. We wanted a clone-and-run path that still respects the upstream model.
What you get
- Docker Compose with NVIDIA GPU passthrough
- PyTorch 2.11 + CUDA 12.8 oriented at RTX 5090 / Blackwell (
sm_120) - Full bfloat16 Miso path by default (~24 GB VRAM recommended)
- Persistent Hugging Face cache volumes
- Preflight checks for gated Llama 3.2 tokenizer access
soundfileaudio I/O patch so PyTorch 2.11 does not depend on torchcodec- Windows
.cmdlaunchers (no PowerShell execution-policy drama) - Web voice demo: Whisper STT, optional LLM replies, dual TTS backends
- Fast mode via Pocket TTS on CPU for low-latency replies
- HTTPS via Caddy for mic access from phones/tablets on the LAN
Two speeds of voice
Quality mode runs Miso 8B on GPU for the high-end voice. Fast mode runs Pocket TTS on CPU so demos stay snappy without monopolising VRAM. You can also run both and A/B engines in the web UI.
Honest requirements
- ≥24 GB VRAM recommended for quality mode
- ~40 GB disk for first-run downloads
- Hugging Face token + accepted Llama 3.2-1B license for the gated tokenizer path
Why it fits HappyMonkey
Voice is becoming a first-class agent interface. Packaging local TTS well is the difference between a LinkedIn demo and something staff can actually use on a desk GPU without a research internship.