Auto-segment long recordings for transcription on low-memory or CPU-only machines #2176

The problem

I'm running Handy on a 2018 MacBook Pro (Intel, 16 GB RAM, no usable GPU), so transcription runs on the CPU. Recordings of around 13 minutes push memory use close to the full 16 GB. The machine stalls and I have to force quit the app, and the recording never gets transcribed.

What I did as a workaround

I had Claude split the saved audio file into shorter segments, update Handy's history database, and show the pieces in the app as separate entries. I then transcribed each segment individually, and that worked fine. Memory use seems to scale with recording length, so splitting a 13-minute file in half keeps usage well within limits. The splitting used ffmpeg to find pauses in the speech, so it didn't cut mid-word or mid-sentence. It works, but it's a manual, clunky process, and I'd rather Handy handle it.

Feature request: automatic segmentation behind the scenes

• Detect available hardware (RAM, CPU vs GPU) and estimate the longest recording it can safely transcribe in one pass.
• If a recording exceeds that, split the audio into chunks automatically, preferably at silences so words and sentences stay intact.
• Transcribe the chunks one at a time, then stitch the results into a single transcript.
• Keep it invisible in the UI: one recording entry and one transcript, with no separate history items for each segment.

Related question: live overlay text vs. the final transcription

I use the live overlay that shows text as I dictate (using Nemotron streaming 3.5 model). When I finish, Handy seems to run a second transcription pass before pasting into the input field, and on long recordings that pass is what fails or stalls. A few questions and ideas:

1. Is the final pass a full re-transcription of the audio, separate from the streaming text? If so, is it there because it's more accurate?
2. If the final pass fails, could Handy fall back to the live transcript it already produced instead of losing everything? Even a less accurate result is far better than nothing.
3. Could the live text be kept or exposed (for example, copyable from the history entry) when the final transcription fails?

Why it matters

Long dictation and brainstorming sessions are where Handy is most valuable, and they're exactly where older hardware breaks down. Automatic chunking and a fallback to the live transcript would make long recordings reliable for anyone without a powerful machine.

Happy to share more details about my setup or help test. Thanks for the great app!