rhapsode — the dictation app I actually use

Creator — open-source macOS app

I use Rhapsode every day, including to write this page. It races fast cloud transcription against a local fallback and checks the audio before trusting suspicious text.

What it is

Why build another dictation app? Honestly, because I wanted one I would actually use. Hold Fn, talk, release, and clean text lands at the cursor in well under a second. I use it all day, every day. This voice profile and this page came through it.

Rhapsode is a signed and notarized macOS app, MIT-licensed, and built on FreeFlow by @zachlatta. I built the resilience stack, learning dictionary, modes, offline pipeline, and voice-bank direction.

The hard part

Can it be fast without losing the dictation when the network dies? Cloud transcription with Groq’s whisper-large-v3-turbo normally lands in about 0.6–0.9 seconds. At the same time, a local whisper.cpp fallback is ready behind a four-second hedge. A bad network costs a few seconds, not the whole thought.

Then there is honesty. Whisper likes inventing a trailing “Thank you.” Did I actually say it? Rhapsode checks the audio energy in that segment’s window. If there was no voice, the text never happened. Real sign-offs survive; hallucinations do not.

What I’d do differently

The fully local path still adds about 1.5 seconds and requires a manual whisper-cpp plus model install. That setup is doing too much. The next move is to make local transcription feel like a normal part of the app, not a side quest.