video-edit-planner-skill/references/windows-funasr-uv-setup.md

Windows FunASR uv setup notes

Use when preparing the bundled transcription project on the Windows host for videos stored on C:\... or D:\....

Execution-location rule

If the video is on the Windows host filesystem, run the long transcription on Windows with PowerShell and Windows paths. Do not give wsl ... transcription commands just because the agent itself runs in WSL. The agent may use powershell.exe from WSL for lightweight setup/debugging, but the user-facing long command should be native PowerShell.

Correct command shape

$TransDir = "$env:USERPROFILE\video-edit-planner-skill\scripts\transcription"
$Video = "D:\Videos\output.mp4"
$VideoDir = Split-Path -Parent $Video

uv run --directory $TransDir funasr-nano $Video --list-tracks
uv run --directory $TransDir funasr-nano $Video --track 2 --language AUDIO_LANGUAGE_CODE --output-dir $VideoDir
uv run --directory $TransDir funasr-nano $Video --track 3 --language AUDIO_LANGUAGE_CODE --output-dir $VideoDir

Outputs should be written next to the video ($VideoDir).

Debugging uv sync from a WSL agent

Capture real stderr/stdout to a log to avoid PowerShell's NativeCommandError wrapping:

$TransDir = "$env:USERPROFILE\video-edit-planner-skill\scripts\transcription"
$Log = Join-Path $TransDir "uv-sync-debug.log"
Set-Location $TransDir
uv sync *> $Log
$Code = $LASTEXITCODE
Write-Host "uv sync exit code: $Code"
Get-Content $Log -Tail 160

From WSL, wrap this in powershell.exe -NoProfile -ExecutionPolicy Bypass -Command '...'.

PyTorch CUDA build pitfall

If funasr-nano logs on cpu on a Windows NVIDIA machine, check PyTorch:

$TransDir = "$env:USERPROFILE\video-edit-planner-skill\scripts\transcription"
Set-Location $TransDir
uv run python -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'no cuda')"

If this prints +cpu, None, and False, the environment has CPU-only PyTorch. Do not fix this with a one-off uv pip install --torch-backend=auto unless the project lock/config is also updated: uv run will sync from uv.lock and may immediately uninstall the CUDA build and reinstall the CPU build.

This project configures torch and torchaudio to use the PyTorch cu130 index on Linux/Windows via [tool.uv.sources]. After pulling a version that includes that configuration, rebuild the environment with:

$TransDir = "$env:USERPROFILE\video-edit-planner-skill\scripts\transcription"
Set-Location $TransDir
git pull
uv sync --reinstall
uv run python -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'no cuda')"

Expected on an NVIDIA Windows host: torch version contains +cu130, torch.version.cuda is not None, and torch.cuda.is_available() is True.

editdistance / Python version pitfall

Observed failure:

Failed to build `editdistance==0.8.1`
error: Microsoft Visual C++ 14.0 or greater is required
hint: editdistance was included because funasr -> editdistance

Cause: the selected Windows Python version may not have a prebuilt wheel for editdistance==0.8.1, so uv attempts a source build. Verified: editdistance==0.8.1 has Windows wheels for CPython 3.12, but not for CPython 3.13.

This transcription project is pinned to Python 3.12 via .python-version and requires-python = ">=3.12,<3.13". If an older clone still selects 3.13, update it and rebuild .venv:

$TransDir = "$env:USERPROFILE\video-edit-planner-skill\scripts\transcription"
Set-Location $TransDir
Set-Content .python-version "3.12"
# Also set pyproject requires-python to a compatible range such as >=3.12,<3.13 if maintaining the project.
Remove-Item -Recurse -Force .venv
uv sync

Avoid defaulting to installing MSVC Build Tools unless the user explicitly prefers compiling native packages on Windows.