Learn · Whisper
Whisper is OpenAI's open-source speech recognition model, and it runs fine on a Mac. The install guides are accurate about the steps and quiet about the evening they can eat. This page lays out both routes: what the terminal path actually involves and when it's the right call, and what to use when you only ever wanted the transcript.
In the left corner · free, some assembly required
DIY
Terminal, models, and real satisfaction at the end
The actual install
$ brew install ffmpeg
$ pip install -U openai-whisper
$ whisper meeting.m4a --model medium
downloading model… 1.42 GBTime to first transcript
An hour if Python, ffmpeg, and the model download all cooperate. An evening if they don’t. whisper.cpp adds a compile step but runs fast on Apple silicon once built.
What you maintain
A Python environment or a compiled binary, ffmpeg, gigabytes of model files, and whatever scripts you write around them. Updates are on you.
Where your audio goes
Nowhere. Local Whisper processes audio entirely on your Mac. If that is your hard requirement, DIY wins this row outright.
Marginal cost
Zero per minute, forever. The real costs are your setup time, disk space, and battery while the model runs.
In the right corner · $8/month
App
Managed OpenAI transcription. No terminal, no models.
The whole setup
Time to first transcript
Minutes. Download Dictate, sign in, hold the hotkey, talk. There is no model picker because there are no models to pick.
What you maintain
Nothing. The model runs on our servers and gets maintained there. Your Mac keeps the app and a signed entitlement ticket, and that’s it.
Where your audio goes
Over TLS to our servers, where it’s transcribed and then discarded. History keeps the text of what you said, never the recording.
Marginal cost
$8 a month flat, no per-minute metering. A subscription you never used is refunded in full.
The first transcript
Here is the part the install guides skip: raw Whisper transcribes audio files. It has no hotkey, no mic capture, and no idea where your cursor is. Finish the DIY setup and you have a very good command-line transcriber, not dictation. Dictate's first transcript lands as words in whatever field you were typing in, because putting them there is the entire product.
New to dictation on a Mac in general? Start with how to dictate on a Mac.
The honest call
If you have a folder of interviews or meeting recordings to transcribe, local Whisper is hard to argue with. Batch jobs run overnight for free, your audio never leaves the machine, and whisper.cpp on Apple silicon chews through files quickly. The same goes if you enjoy the tinkering itself. Some people want a working pipeline they built, and a Whisper install is a satisfying weekend.
Where it stops making sense is live dictation. Nobody wants to pipe their mic through a Python script forty times a day. For that you want an app, and the question becomes which kind.
Middle path · local files
Whisper models in a native Mac app, pointed at files. Drag in a recording, get a transcript, no terminal involved. It's built for transcribing recordings more than for live dictation. We compare the two jobs in MacWhisper alternatives for live dictation.
Middle path · local dictation
Local models with a real dictation loop: hotkey, speak, insert. You still pick and download models, but it's a Mac app, not a script. If local processing matters to you and DIY doesn't, it's the pick to evaluate. Our side-by-side is at Dictate vs VoiceInk.
Superwhisper sits in this camp too, with a model picker of its own. If you're headed that way, read Superwhisper alternatives first. Dictate is our app.
Setup
No Python, no ffmpeg, no 3 GB model downloads, no picking between tiny and large. You sign in and the accuracy is simply what our servers run, kept current without you noticing.
Insertion
Hold the hotkey, speak, release. The transcript is inserted into the field you were focused on, in any app. That insertion step is exactly what raw Whisper doesn't do.
The trade
Over TLS to our servers, transcribed, then discarded when the request completes. History keeps text only, never recordings. If that trade is unacceptable, stay local: run whisper.cpp or a local app like VoiceInk instead.
Whisper on a Mac, answered
Setup, cost, and local processing.
Yes. Install openai-whisper with Python and ffmpeg, or build whisper.cpp natively on Apple silicon. Download a model separately, then transcribe audio files from the command line.
No. OpenAI released code and model weights, not a desktop app. Third-party apps such as MacWhisper and VoiceInk package it locally. Dictate uses managed server-side transcription.
The code and weights are free. You still spend setup time, disk space, and local compute. Packaged apps charge for the interface; Dictate costs $8 a month.
Not by itself. Whisper transcribes audio files but does not provide a hotkey, microphone capture, or cursor insertion. A dictation app adds those pieces.
No. Dictate sends audio over TLS for server-side transcription, then discards it. History keeps text, not recordings. For fully local processing, use whisper.cpp or a local app.
Managed OpenAI transcription, $8 a month flat. Cancel any time; a subscription you never used is refunded in full. Details on the one-plan pricing page.