Fully offline transcription on a Mac is possible because Apple's speech models run natively on Apple silicon: after macOS downloads a small model for your language (once), TranscribeAnything transcribes with the network cable unplugged — no account, no upload, no server. That makes it suitable for recordings that must not leave the machine: privileged legal interviews, patient sessions, unreleased product discussions, journalism sources.
Most transcription products describe themselves as private while still uploading your audio to their servers for processing. On-device transcription is architecturally different: the speech model is stored on your Mac, inference runs on your Mac's Neural Engine, and the resulting text is written to your disk. There is no server component — nothing to breach, subpoena, or train on.
The two exceptions are explicit: fetching a web link requires the network (the file has to come from somewhere), and checking for app updates does. Neither involves your audio.
Lawyers with privileged recordings, clinicians and therapists with session audio, journalists protecting sources, researchers under IRB constraints, and companies whose security policy simply prohibits third-party processors. In each case the compliance conversation gets dramatically shorter when the honest answer to 'where is the audio processed?' is 'it never leaves the laptop.'
TranscribeAnything runs Apple's on-device speech engine at up to 80× realtime, supports 10 languages, and costs $19.99/year for unlimited transcription — the first 5 transcriptions (up to 30 minutes each) are free.
Apple's speech engine is the same technology that powers system-wide live captions and dictation across macOS — modern, state-of-the-art on-device recognition. On clean audio it is comparable to cloud services; the only thing you give up is someone else's computer.