Prefer DirectML for Windows ONNX transcription models

[ferologics]

Summary

• patch Handy's transcribe-rs dependency to a forked git revision with Windows DirectML support for ONNX models
• prefer DirectMLExecutionProvider on Windows, with explicit CPU fallback if provider registration fails
• log whether DirectML registration succeeded or fell back to CPU
• clean a few existing Rust warnings touched during validation

Validation

• cargo check
• cargo check --release
• launched the built Handy app locally and confirmed handy.log shows successful DirectML registration for the Parakeet ONNX sessions
• ran a local dev-build transcription successfully after the change
• benchmarked the patched Parakeet path on a real Handy recording (244.38s audio): 35.368s before vs 6.99s after (~5.1x faster, ~35x realtime)

Dependency patch

• Handy now patches transcribe-rs to this git revision:
  • ferologics/transcribe-rs@c56480687127070f456ae462d73c5defe964d807
• Upstream PR for the current transcribe-rs mainline:
  • Prefer DirectML for Windows ONNX sessions transcribe-rs#53

[github-actions]

🧪 Test Build Ready

Build artifacts for PR #985 are available for testing.

Download artifacts from workflow run

Artifacts expire after 30 days.

[cjpais]

@ferologics can you see if this helps the inference speed for you still, would be quite curious if it just goes back to CPU or works out the box. I kind of think since it's direct ML it might just work on Win 11, would be curious about Win10 too

[ferologics]

Tested the CI-built Windows artifact locally and it looks good.

What I checked:

• downloaded handy-pr-985-x86_64-pc-windows-msvc from run 22882036165
• extracted the MSI payload and confirmed the packaged app includes DirectML.dll
• launched the packaged handy.exe
• triggered a real start/stop transcription cycle against my normal Handy setup

Result:

• handy.log shows successful DirectML registration in the packaged build:
  • ONNX Runtime session registered DirectMLExecutionProvider on Windows (device 0) with CPU fallback enabled
• I saw that log for the Parakeet encoder / decoder / nemo sessions
• I did not see the CPU fallback warning

So on my Windows 11 machine the CI-built artifact is still taking the intended GPU path, not silently dropping back to CPU.

[cjpais]

Solid. This is amazing news. I will test on my windows machine when I can and see how it goes as well. I'm curious how this will play with integrated GPUs

I am slightly wondering if we will need to provide an option to disable this. Just in case CPU is faster for someone. I know another PR in transcribe rs had something like this. Might be worth considering