Whisper models

Choosing a model

Engine	Download	Character
Tiny	39 MB	Fastest. Fine for quick notes; struggles with accents, noise, and jargon.
Small	466 MB	Everyday dictation on any Mac. The safe default on Intel.
Medium	1.5 GB	Hears more, asks more of your Mac.
Large turbo	1.6 GB	Hears the most. Best on Apple Silicon, and fast enough for live dictation.

During onboarding one option carries a RECOMMENDED chip, picked from your actual hardware: Large turbo if you are on Apple Silicon with more than about 5 GB free, otherwise Small. That heuristic is a good default — the sections below cover when to override it.

When to go bigger

• You dictate in a language other than English — model size matters most here.
• You have an accent the smaller models handle poorly.
• You dictate in rooms with background noise or at a distance from the mic.
• Your work is full of technical vocabulary.

When to go smaller

• You are on an Intel Mac, where there is no Neural Engine to lean on.
• Your Mac has 8 GB of memory and is already under pressure.
• You want the shortest possible delay between releasing the key and seeing text.
• Disk space is tight.

Changing your model

Go to Dashboard ▸ Settings ▸ Transcription — on this Mac ▸ Whisper model and pick one. If it has not been downloaded before, the download begins immediately; models you have used before are already on disk and switch instantly.

Downloads and storage

Models download once and are cached at ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/. The progress display shows transferred size, current speed, and a time estimate.

Two behaviours worth knowing:

• The bar reaching the top does not mean it is done. The transfer occupies most of the bar, then macOS compiles the model for your Mac — labelled “now optimizing for your Mac. First run only.”
• Switching engines mid-download cancels the transfer in progress rather than running two at once.

If a download fails

The app retries once automatically, deleting the partial files first — a truncated download is the usual cause of a model that will not load. If the second attempt also fails, the screen shows Try again plus a Details disclosure with the underlying error.

The app deliberately does not guess at causes. If Details mentions the network, check your connection or VPN; if it mentions disk space, free some and retry. Manual recovery: quit the app, delete the model folder above, relaunch, and select the model again.

Frequently asked

Which Whisper model should I use?

On an Apple Silicon Mac with a few gigabytes free, use Large turbo — it is the most accurate option that still feels immediate. On an Intel Mac or a machine short on disk or memory, use Small. Tiny is for quick notes where speed matters more than precision.

Where does OpenVoiceFlow store downloaded models?

In the WhisperKit cache under your Documents folder, at ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/. Deleting that folder frees the space; the app re-downloads the model you select next time.

Why does the progress bar pause near the end of a model download?

The transfer finishes around ninety percent, then macOS compiles the model for your specific Mac. That compile step happens only on the first run of a given model and is why the app shows 'Downloaded — now optimizing for your Mac'.