OpenWhisper 0.5 transcribes as you talk and spells your jargon right
3 hours ago
OpenWhisper is a menu bar app that turns your voice into text and pastes it into whatever app you're using, all on your Mac. Version 0.5 adds models that transcribe while you're still talking, and a vocabulary that gets names and jargon right. Apple's built-in engine, though, needs macOS 26.
Ask about this presentation
Chapters
Show transcriptHide transcript
OpenWhisper is a menu bar app that turns your voice into text and pastes it into whatever app you're using, all on your Mac. Version 0.5 adds models that transcribe while you're still talking, and a vocabulary that gets names and jargon right. Apple's built-in engine, though, needs macOS 26.
Slide text
OpenWhisper 0.5 transcribes as you talk and spells your jargon right
9 streaming models · a Vocabulary tab · Press & HoldOpenWhisper 0.5.0The four Whisper models wait until you stop, then transcribe the whole recording. The new streaming models transcribe the audio as it arrives, and the main window's status line shows the latest words while you speak. When you stop, the app finishes that same stream.
Settings now has a searchable Voice Model picker. Apple's engine sits on top, then eight downloadable streaming models, then Whisper, which adds a Large option at 1.1 gigabytes. Parakeet Unified scores highest on accuracy, and Moonshine Tiny is the smallest and fastest, at 50 megabytes.
A new Vocabulary tab holds the names and terms you use, on top of about 230 built-in ones like PostgreSQL and OAuth. Your list stays on this Mac. Whisper and Apple's engine get those terms as a hint. Then a local pass corrects close misses in the final text, whichever engine you picked.
If you say PostgreSQL and the model hears "poster SQL", you still get PostgreSQL. A garbled NYSE gets fixed too. Add AcmeDB to your list, and a lowercase acmedb comes out spelled your way. Ordinary prose is left alone, so Jason stays Jason.
A new Recording Mode, Press and Hold, records while you hold the hotkey and transcribes when you let go. Toggle stays the default. You can now cancel while the app is processing. Cleanup removes only sound markers like blank audio, so a TODO in brackets survives. And if Accessibility access is off, the app says so, and your text waits in History.
Three of the four pull requests in this release are about testing and review. Every pull request now feeds recorded audio through Whisper. Silence has to come out empty, noise nearly empty, and a short English clip has to come back with its key phrases. Alongside that come unit tests, UI tests that click through the app, and automated code and security review.
Nothing here is marked as breaking, and an existing install keeps its Whisper model. Before you switch, know that Apple's engine needs macOS 26 and may download its speech model the first time. Six of the new models are English only. Voxtral is a 2.8 gigabyte download. And Multitalker Parakeet doesn't label speakers.
To upgrade, choose Check for Updates in the menu bar, or download the disk image from the Releases page. Then try a streaming model on your own voice, because those speed scores came from a PC, and add your team's names to the Vocabulary tab. That's OpenWhisper 0.5, transcribing as you talk and spelling your jargon right.