Whisper for Mac

Why run Whisper locally

OpenAI's Whisper is the speech recognition model that made high-quality transcription accessible to everyone. You can call it through the OpenAI API — that's the path most people take first. But the model weights are open, and on Apple Silicon you can run them locally, at native speeds, without your audio ever leaving your Mac.

There are three reasons to do this. One: privacy. Everything stays on your device — meetings, memos, sensitive conversations, the stuff you wouldn't put in a third-party tool. Two: cost. The API charges per minute; local inference on hardware you already own is free after the initial model download. Three: latency and offline. No network round-trip means faster dictation, and it works on airplanes, in air-gapped environments, anywhere you don't have a connection.

The three paths to Whisper on Mac

1. whisper.cpp (command line)

whisper.cppis Georgi Gerganov's ultra-fast C++ port of Whisper. It runs natively on Apple Silicon with Metal and Core ML acceleration, supports every Whisper model size (tiny to large-v3 and Turbo), and can transcribe audio files directly from the terminal. It's what most Mac-native Whisper tools are built on under the hood.

Good for: developers, scripted pipelines, file transcription, local servers. Not a dictation experience on its own — you still need something on top that handles hotkeys, audio capture, and pasting into the focused text field.

2. Python / OpenAI whisper (reference implementation)

The original OpenAI whisperPython package works on Mac too. It's a cleaner codebase for research and experimentation, with access to every official model variant. On Apple Silicon it's slower than whisper.cpp because it doesn't use Core ML or Metal as aggressively, but it's the reference for checking behavior and model output.

3. A native Mac app built on local on-device STT

Most people don't want to run a Python script to dictate an email. They want a hotkey that pops open a mic and pastes clean text where the cursor is. SuperWhisper, VoiceInk, and MacWhisper are Mac apps built on local Whisper for exactly this reason — they wrap Whisper in an actual product.

Resonant is in the same category — a free Mac app for live, hotkey-driven dictation — but built on a different model. It runs Parakeet (NVIDIA's on-device STT) on Apple Silicon's Neural Engine, with filler-word removal, sentence cleanup, and custom vocabulary on top of the raw transcription. If you specifically want Whisper-architecture transcription, whisper.cpp is still the standard route, and MacWhisper, VoiceInk, and SuperWhisper all wrap Whisper in different ways.

Which Whisper model should you use on Mac?

Whisper comes in several sizes — tiny, base, small, medium, large-v2, large-v3 — plus Turbo and distilled variants. Bigger models are more accurate but slower. For live dictation on Apple Silicon, Turbo and distil-large hit the best point on the speed/accuracy curve for most English use cases. For multilingual accuracy, large-v3 is still the gold standard.

Resonant doesn't use Whisper — it ships Parakeet, tuned for live dictation latency on Apple Silicon. If you're building your own Whisper pipeline, Turbo is usually the right starting point on M-series chips.

Core ML and the Neural Engine

Apple Silicon includes the Neural Engine — a dedicated co-processor for machine learning workloads. Whisper models can be converted to Core ML format to run on it, which is dramatically more power-efficient than CPU or GPU inference. whisper.cpp supports Core ML acceleration, and well-built Mac apps take advantage of it to keep dictation fast without burning battery.

The easy answer

If you want to run Whisper on Mac and you already know why — whisper.cpp is the right foundation and it's free. If you want the dictation experience — press a key, speak, clean text in any app — download Resonant. It's built on Parakeet rather than Whisper, but solves the same problem: turning your voice into clean text in any macOS app, fully on-device.

Frequently asked questions