Whisper dictation on a Mac

Whisper is OpenAI’s open speech-recognition model. A “Whisper app for Mac” is an app that runs one of its checkpoints — on your Mac, or on someone’s server. Which checkpoint, run where, decides almost everything you will notice: accuracy, speed, the download, and whether your voice leaves the machine.

Three ways to get Whisper on a Mac

• A dictation app — hold a key, speak, and the text lands wherever you were typing. This is what most people searching for a Whisper app for Mac want, and it is what Vocula is. superwhisper and VoiceInk are the other well-known ones.
• A transcription app — you give it a recording, a meeting or a video and it gives you a transcript, often with timestamps and speakers. MacWhisper is the best-known, and for files it is the better tool.
• The command line — brew install whisper-cpp gives you whisper-cli, the same engine Vocula is built on. Free, and fine for a folder of recordings; nobody dictates an e-mail this way.

What Vocula runs, exactly

Every line below is in the app’s source, in one file you can read.

“Turbo” is large-v3 with its decoder cut from 32 layers to 4. The part that listens is unchanged; the part that writes is much smaller, which is most of why it is fast enough to feel instant on a laptop.

Five questions to ask any Whisper app

1. Does the audio go to OpenAI’s API? “Powered by Whisper” can mean a model on your Mac or an upload to OpenAI’s servers. Both are Whisper. Only one is private.
2. Which checkpoint? tiny, base, small, medium, large-v3 and turbo are different products. The small ones are quick and noticeably worse at names, accents and anything that is not English.
3. Where does the model come from? Some apps fetch it from a third-party model host at first launch. That is a network dependency you did not choose, and a file that can change under you unless it is pinned by hash.
4. GPU or CPU? The same checkpoint on the CPU is noticeably slower. On Apple Silicon, Metal is the difference between waiting and not noticing.
5. What happens in silence? Whisper was trained on subtitled video, and given silence it will sometimes invent a line — “Thank you for watching” is the famous one. An app needs to find the speech before Whisper hears it, and throw away output that is only a [Music] tag. Vocula does both.

How to check the first one yourself, on any app, in ten minutes.

Where Whisper is weak, including in Vocula

• It can pick the wrong language. Whisper decides the language from the start of each recording. Vocula only lets it choose among the languages you selected, which rules out the ones you never speak — but it can still pick the wrong one of yours, and a very short or unclear phrase is where that happens.
• No live preview. Vocula gives you the text when you let go of the key, not word by word while you talk.
• 1.62 GB to download, once. If that is too much, or all your languages are European, Vocula’s second engine is Parakeet: 468 MB and faster, 25 languages.
• Apple Silicon and macOS 26 only. No Intel Macs, no older macOS.