Speech to text on a Mac.

Three jobs, one idea

People say speech to text, voice to text, dictation, and transcription and mostly mean one of three jobs. Live dictation: you talk and text appears where your cursor is, in real time. File transcription: you already have the audio (a lecture, a meeting recording, a voice memo) and want it as text. And voice control, which is about commanding the machine rather than writing, and is a different tool entirely (on a Mac, that’s the Voice Control accessibility feature, not Dictation).

The reason the distinction matters: the tools are shaped for one job each. Picking a meeting-bot to dictate into a terminal, or a dictation app to transcribe two hours of recorded audio it wasn’t built to ingest, is how most speech-to-text disappointment happens.

The built-in route

macOS ships Dictation free: System Settings, then Keyboard, turn on Dictation, then press the shortcut in any text field and talk. Apple’s guide notes it stops automatically after 30 seconds of silence, and that Keyboard settings show whether your speech is "processed on your device and not sent to Siri servers", which it can be for supported languages. It has no custom vocabulary and thin command support, but it costs nothing and it’s already installed.

The full walkthrough on this site covers enabling it, the punctuation commands, the usual fixes when it stops working, and its real limits. If you dictate occasionally, start there; you may not need anything else.

The apps, and where your speech is processed

Every claim below is from the vendor’s own site, checked on the date at the top of this page. Price models rather than prices, since prices move.

On-device or cloud: how to actually tell

On-device means the audio is turned into text by models running on your Mac, and the recording never crosses the network. Cloud means your voice is sent to a server, processed there, and the text comes back. Cloud engines are a legitimate choice, often stronger on accents or older Intel Macs, and Wispr Flow says plainly that cloud is how it gets its speed and accuracy. The problem is only ever ambiguity.

So don’t take an adjective’s word for it. "Private" and "secure" describe policies; on-device describes an architecture. The vendors doing local processing say so in checkable terms: VoiceInk’s "your voice data never leaves your Mac" is a falsifiable claim from an open-source codebase, Spokenly ships a mode that blocks network requests, and Apple tells you in System Settings which side of the line your language is on. If a product page won’t say where speech is processed, that is itself the answer.

Recorded audio: transcription, not dictation

For files, the question changes: hours of audio instead of seconds, and speaker labels start to matter. MacWhisper is the best-known Mac tool here and earns it, transcribing locally with a free tier that covers a lot of use. Edict covers the same ground its own way: drop an audio or video file on the app and it transcribes on your Mac with speaker labels, exporting SRT, VTT, or plain text, and your agent can submit files over MCP and get the transcript back without any audio leaving the machine. A confidential interview, a lecture recording, a meeting you were sent: none of it needs to touch a server.

What you can do for free

You can do real speech-to-text work on a Mac for nothing: Apple Dictation is free and already there. Spokenly’s local transcription is free. MacWhisper’s free tier transcribes files. superwhisper has a perpetual free plan. VoiceInk is open source, so you can build it yourself. Paid apps have to earn the difference in daily-driver details: vocabulary that spells your jargon right, no silence timeout, transcript history, guardrails for agent work. If the free routes cover you, use them; this site will not pretend otherwise.

Where Edict fits

Edict is the always-on-device end of the spectrum: one key, speech models inside the app so the first launch works offline, verbatim output with your own dictionary applied, a wrong-window guard for agent work, file transcription with speaker labels, and an MCP server so your agent can use all of it. It dictates in 25 European languages and detects which one you are speaking without being told, though its own interface is English only. One-time purchase, 7-day full trial, macOS on Apple Silicon. If you dictate into coding agents all day, it was built for exactly that; if you mostly need free-form notes twice a week, the built-in tool above is fine, and the comparisons will point you elsewhere when a competitor fits better.

Common questions