Local Transcription

What that means in practice#

No round trip. Speed is bounded by your Mac, not by your connection or by someone else's queue. A short phrase does not wait behind a network request.

Offline capable. The speech model is downloaded once, on first run. After it is installed, dictation works with no internet at all. Purchases, model downloads and update checks still use the network when they need to.

Predictable at rest. The menu-bar agent stays out of the way with near-zero CPU while idle, and keeps the selected whisper.cpp context resident after warm-up so the next dictation starts quickly. Memory use scales with the model you chose.

Choosing accuracy against weight#

Large v3 Turbo is the recommended default. Base, Small and Medium are available when you prefer a lighter model, whether for memory headroom or for speed on a busy machine.

This is the main knob you have over the transcription state. Choose a Speech Model covers the trade in detail.

The limits worth knowing#

The speech model accepts a limited number of custom dictionary terms before it ignores the rest. Phantom Voice used to send the whole list silently; current builds mark any term that does not fit as not sent in Preferences, so you know which of your terms are actually active. See Personal Dictionary.

English only, on macOS 14 or later, on Apple silicon. Full detail in Requirements.

Next#

• Privacy on Device — what this architecture means for your data
• Optional Cleanup — the second local model, and why it is checked
• Back to the documentation home