Optional remote or local server for the transcription step

[zft9xgy]

First, a thank you. I use Handy every day on my Mac and it works great. I have donated and I really appreciate the time and effort that goes into building and maintaining it.

I would like to suggest an optional feature that could help people with lower-powered devices, and also people like me when the machine is already under heavy load.

The idea

Let the transcription step send the recorded audio to an external endpoint, instead of always running it on the local device. Post-processing already supports something like this, since it can point to an Ollama instance that is not local. The transcription step does not offer an equivalent option, and that is the gap I would like to close. This would be opt-in and off by default, so nothing changes for anyone who wants the current fully local behavior.

The problem I run into

With heavier models, the local run can struggle on long recordings. In my case, using the Cohere Transcribe model, audio around 5 minutes sometimes does not finish transcribing. It looks like it reaches a memory ceiling and never returns a result. While it processes a long clip, the machine also lags.

Why this does not break the offline spirit

The endpoint does not have to be a cloud service. It can be your own machine on your local network, or a self-hosted box in a homelab, just like a remote Ollama instance. In that case the audio never leaves your control, and you simply move the heavy work off the device you are dictating on. A public endpoint would just be one more option for those who want it.

Possible shape

For the transcription step, a minimal setting could mirror what post-processing already allows:

• A toggle to enable remote transcription
• An endpoint URL
• An optional auth token
• Fallback to local processing if the endpoint is unreachable

Happy to help test this on macOS. Thanks again for the project.