Background:
My laptop was stolen recently and the new one I got to replace it has an Intel NPU [0] in it. The promise of the NPU is running small machine learning models efficiently on mobile hardware. I thought a good application of this would be using whisper to transcribe speech into text. There’s not really much [1] out there on Linux that can do this right now which is kind of a bummer because it’s a big accessibility thing to be able to type with your voice. I use my Sway configuration [2] to map the right control key to run a wrapper program [3] and then the release of the right control key to send a SIGINT to that program. The wrapper catches the SIGINT, ends transcription, and types the transcribed text into the focused application with the `enigo` crate.
Repo link: https://github.com/ellenhp/whisper-npu-server
This is not one of my high polish projects, but I did want to throw it out there into the world, especially because the OpenVINO project doesn't have any containerized NPU examples, even for LLMs.
[0] https://intel.github.io/intel-npu-acceleration-library/npu.h...
[1] I found this, and based some of my code on it: https://github.com/oddlama/whisper-overlay
[2] See end of post for example.
[3] https://github.com/ellenhp/whisper-transcription-wayland/
Sample Sway config:
bindsym --no-repeat Control_R exec "whisper-transcription"
bindsym --release Control_R exec killall -2 whisper-transcription
