Guide

How to run Whisper locally on a Mac

OpenAI's speech model is open source, so it can run entirely on your own machine. Three ways to do it, from a Terminal command to an app that needs no setup, and how to pick a model size.

The short answer

There are three ways to run OpenAI's Whisper on a Mac without uploading anything: the original Python package, whisper.cpp, or an app that has whisper.cpp built in.

For files on Apple Silicon, whisper.cpp is the one to use from Terminal: it is fast and uses the Mac's GPU. If you want speech typed into other apps as you talk, the command line is the wrong tool and an app is the right one.

What Whisper is, and what "locally" means

Whisper is a speech recognition model that OpenAI released as open source in 2022, with its code and model weights published under the MIT licence. Because the weights are public, anyone can run it on their own hardware — which is the whole point here. When it runs locally, your audio is read from your disk, processed by your Mac's own chip, and the text is written back to your disk. No account, no API key, no per-minute charge, and it works with the Wi-Fi off.

That is different from OpenAI's paid transcription API, which also offers Whisper but processes your audio on OpenAI's servers. Same model family, opposite privacy property.

Choosing a model size

Whisper comes in sizes. Bigger models are more accurate, especially on accents, jargon and noisy recordings, and are slower and larger on disk. The disk sizes below are whisper.cpp's model files.

Whisper model sizes, as published by OpenAI and whisper.cpp
ModelParametersFile on diskGood for
tiny39M~75 MiBTesting that everything works
base74M~142 MiBFast dictation of clear speech
small244M~466 MiBA balanced default for files
medium769M~1.5 GiBHarder audio, more patience
large1550M~2.9 GiBThe most accurate, the slowest

Two variants are worth knowing. Models ending in .en are English-only and a little more accurate on English at the smaller sizes; use the plain names if you speak anything else. large-v3-turbo is a slimmed-down large model that is much faster than large for a modest loss in accuracy, and is a popular choice for long recordings on Apple Silicon.

Way 1: the original Python package

This is OpenAI's own reference implementation. It is the easiest to install if you already use Python, and the slowest on a Mac, because out of the box it runs on the CPU rather than the GPU.

  • Install ffmpeg, which Whisper uses to read audio: brew install ffmpeg
  • Install Whisper: pip install -U openai-whisper
  • Transcribe a file: whisper recording.m4a --model small --language en

The model is downloaded the first time you use it and cached. By default it writes the transcript in several formats at once — plain text, SRT and VTT subtitles, TSV and JSON — next to where you ran it. Leaving out --language makes Whisper detect it from the first thirty seconds.

Way 2: whisper.cpp, the fast one

whisper.cpp is a C/C++ reimplementation of Whisper by the ggml project. On Apple Silicon it uses the GPU through Metal, which makes it much faster than the Python package on the same Mac. You need Apple's command line tools (xcode-select --install) and CMake (brew install cmake).

  • Get the code: git clone https://github.com/ggml-org/whisper.cpp and cd whisper.cpp
  • Download a model: sh ./models/download-ggml-model.sh base.en
  • Build it: cmake -B build then cmake --build build -j --config Release
  • Test it on the bundled sample: ./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav

whisper.cpp is happiest with 16 kHz mono WAV. To be safe, convert anything else first: ffmpeg -i interview.m4a -ar 16000 -ac 1 -c:a pcm_s16le interview.wav. Add -otxt or -osrt to the whisper-cli command to save a text or subtitle file instead of only printing to the terminal.

Way 3: an app with Whisper inside

Both command-line routes are good at one job: turning a file into text. They do not do the thing most people actually want a voice tool for, which is speaking into whatever app they are already in. For that you need something that records from the microphone on a shortcut, runs the model, and types the result at your cursor — and several Mac apps are exactly that, built on whisper.cpp.

Voz is one of them. It bundles whisper.cpp and the Whisper base model inside a single 146 MB download, so there is nothing to build, download or configure: install it, allow the microphone and Accessibility (which lets it paste the text for you), press your shortcut and talk, press it again and the text appears where your cursor is, in any app. The same local model also transcribes meetings recorded on your Mac and audio or video files you drop in, and keeps them in a searchable history. Nothing is uploaded; there is no account. It costs $2.99 once.

One detail for people who have already done Way 2: Voz looks in ~/whisper.cpp/models. If ggml-small.bin, ggml-base.en.bin or ggml-large-v3-turbo.bin is there, Voz uses it by default instead of its bundled base model (checked in that order), and other ggml-*.bin files there appear in its model menu. Note that base.en is English-only, so if that is the one it picks up and you dictate in another language, choose a multilingual model in Settings. Meetings have their own model setting, separate from quick dictation.

Other apps in this space include MacWhisper, built for transcribing files with timestamps and subtitle export, and Superwhisper and VoiceInk, which type at your cursor like Voz with more settings. Their prices are in our dated table of every Mac dictation app.

Which route to take

  • You want subtitles or timestamps for recordings and like Terminal: whisper.cpp, with -osrt.
  • You already live in Python and speed does not matter: the openai-whisper package.
  • You want to talk instead of type, in email, chat, documents or a code editor: an app. The command line cannot type into other windows for you.
  • You need the best possible accuracy on difficult audio: a large or large-v3-turbo model, through whisper.cpp or an app that lets you choose it.

To be fair to the free routes

Both command-line options are free and open source, and for transcribing files they are excellent. If that is all you need and you are comfortable in Terminal, you do not need to buy anything — Voz included. An app earns its price by removing the setup and by typing where you work.

Questions people ask next

Does local Whisper need an internet connection?

Only to download the code and the model the first time. After that, transcription runs with no connection at all.

Will it run on an Intel Mac?

Yes, but slowly. Apple Silicon is where whisper.cpp's GPU acceleration makes the difference; on Intel, stick to the smaller models.

Is it as accurate as cloud transcription?

The large models are very accurate on clear speech, though how they compare with a given cloud service depends on the audio and the service. The small models trade accuracy for speed, and you will notice it on accents, jargon and noisy rooms.

Can Whisper learn my vocabulary?

Not by training, but it accepts a short text prompt that steers its spelling of names and jargon. Here is how custom vocabulary works.

Is running Whisper locally private?

Yes. The audio never leaves your Mac. The only data that crosses the network is the code and model you download.

Sources: OpenAI's Whisper repository and model card for model sizes and licence; the whisper.cpp repository for build steps and model file sizes. Voz details from the shipping app's source code.

Voz is an independent product and is not affiliated with, endorsed by or sponsored by OpenAI or any other product named on this page. Product names and trademarks belong to their respective owners. Commands change between releases; if a step here no longer works, tell us — we would rather correct this page than defend it.