Mac dictation that types as you talk, and meeting notes without a bot
WunderWhisper is a macOS menu bar app I built and sell myself. Hold Option+Space and your words get typed into whatever app has the cursor while you talk, then a small on-device model cleans them up when you let go. The Pro side records meetings from your microphone and the call audio, so no bot joins the call, and writes the transcript, speaker labels and notes on the same Mac. Speech recognition never leaves the machine. Dictation is free with no word cap, and Pro is $29 once.
Key information about the project scope and deliverables
Wunderlandmedia
Productivity Software / macOS
1 week to the first paid version
2026
How we transformed obstacles into opportunities
Apple's speech engine only settles a sentence at a pause. One 7.6 second sentence came through at the very end, so typing finished text meant nothing appeared until you stopped talking.
AI cleanup can flip a sentence by dropping a "not" or a "never". I found that complaint about a competing app while researching this one.
Replacing text inside someone else's app is risky. Terminals pass every Accessibility check while the shell line can't be selected, and Chromium apps ignore the write, so a careless swap deletes the wrong characters.
A meeting can't be recorded twice. A crash, a headset unplugged mid-call or a Mac going to sleep can't cost you the call.
It had to work without installing a second app, which meant one small model good enough for instant dictation cleanup and for notes on an hour-long meeting.
Provisional words are typed about a second after you say them, holding back the last word, and the settled text swaps in when you let go.
An output guard counts negations in English and German. If the count changes, the cleanup is thrown away and your own words stay.
The clean version replaces your dictation only if Accessibility shows exactly that text right before the caret. Terminals keep the raw text and Copy Last Dictation holds the clean one.
Transcript lines hit the disk as they arrive, the audio format survived a kill -9 in testing, capture restarts on device changes, and an interrupted meeting can be processed again.
I ran 15 models through cleanup tests and a scripted one-hour meeting. Qwen3 4B became the default: 0.4 seconds median cleanup, meeting notes as good as the 8 to 9B models, 2.3 GB on disk.
Workflow
A systematic approach to delivering exceptional results
I started from my WunderType app, kept the AI provider layer, Keychain and settings, and dropped the rest. Scope: meetings and dictation in one menu bar app, shipped outside the App Store.
Microphone and call audio go through one 16 kHz format into Apple's SpeechAnalyzer, FluidAudio handles speaker labels, and MLX runs the cleanup and notes on the Mac's GPU.
A scripted meeting with planted decisions and traps, typing tests against a separate helper app, and 147 automated tests by the latest release. Bugs from my own use went straight into the next build.
Signed and notarized builds, Sparkle updates, free dictation plus a $29 Pro tier on Polar license keys with no license server of my own, and an Astro site at wunderwhisper.com.
Visual showcase of the project's key features and design elements