Private On‑Device Dictation

What stays local

On‑device models run entirely on your Mac. Audio is processed locally, and transcripts are stored locally in History. You control retention—keep, search, clear, or export.

On‑device models

Choose Whisper sizes that balance accuracy and resource use, then download and activate them. With Apple Silicon and GPU acceleration, you get practical, low‑latency dictation without sending audio to a server.

Local storage, predictable behavior

Notes are saved in History by default. Browse and search instantly, and copy or share any version (Original, Replaced, LLM). You can clear history, disable auto‑save, or export specific entries.

Cloud when you need it—explicit

Cloud speech‑to‑text and LLM features contact servers only when enabled. When a cloud model is selected, only the audio required for that request is sent. Switch back to on‑device to keep everything local.

Accuracy and workflow

On‑device Whisper handles long utterances, domain terminology, and multilingual speech. If the LLM takes too long, cancel and dictop falls back to the Replaced version.

Permissions

• Microphone: required for recording.
• Accessibility (macOS): needed for paste/typing insertion and muting speakers.
• Screen Recording (macOS): needed to capture text for LLM placeholders. Allowed Apps controls where capture is permitted.

Defaults

Out of the box: On‑Demand mode; copy output rather than typing; Insert Latest prefers LLM, then Replaced, then Original; prepend one space; no append; History on; sounds on; pause shortcut uses press/release; Language Auto; silence threshold 3 seconds; Whisper Tiny active by default.

Privacy notes

On‑device models keep both audio and text local. When you use LLM placeholders for screen or app text, extraction happens locally and only the text string is included in your prompt.