Windows and NVIDIA deployment notes

This path is a recommendation for the planned dedicated Windows/NVIDIA host. It has not yet passed the Local Voice acceptance suite and must be treated as provisional.

Recommended host

Windows 11
Current NVIDIA Studio or production driver
Python 3.11 x64
ffmpeg and Whisper available on PATH
Sufficient SSD space for the 4.7 GB model, jobs, and retained WAV files
NVIDIA GPU with supported CUDA capability and practical VRAM headroom

Preferred deployment order

Start with native Windows and a CosyVoice3 wheel built for the installed CUDA runtime.
If native dependency resolution is unreliable, use WSL2 with NVIDIA CUDA passthrough and the Linux CUDA build.
Use LOCAL_VOICE_RUNTIME to point at a dedicated data directory, for example D:\MPM-Local-Voice\runtime.
Copy the same private voices archive and model directory used on macOS.
Run doctor --device cuda.
Run the full acceptance suite before scheduling production briefings.

Important differences

The verified Metal wheel cannot run on Windows.
CUDA, driver, and wheel versions must agree.
Do not assume that a CUDA build exists merely because the NVIDIA driver is installed.
Keep ffmpeg path quoting and Windows long-path behavior in mind.
Compare transcript coverage, ending confidence, seam derivatives, duration, and subjective voice similarity against the macOS acceptance outputs.

Fallback

CPU generation remains functionally possible but may be slower. It is suitable for overnight batches if CUDA setup is delayed, provided the acceptance suite passes on that host.

Certification checklist

cosyvoice3, NumPy, and SoundFile import
Candle model hashes match the recovery manifest
All authorized voice assets resolve
CUDA device loads successfully
Every acceptance output passes transcript QA
Ryan fixed assets and dialogue assembly are seamless
Five production briefs can complete inside the overnight window

2.0 KiB Raw Permalink Blame History