Just Speak to It
Download
Hold a key and talk, and the words land in whatever you were typing in. You pick the speech model — one that runs on your own Mac, or a cloud provider you hold the key to.
Could we move the catch-up to Friday?
Native Mac, iPhone and iPad On-device or your own API keys No account, no subscription Read the source
How it works
No window to find, no button to hunt for. The app stays out of the way until you hold the key.
Press and hold Fn, double-tap it, or record your own global shortcut — the app answers to all three out of the box. It fires while another app is frontmost, so Just Speak to It is never the window you are looking at.
fn hold, double-tap, or bind your own
Words appear as you speak them, or hand over a recording and get it back transcribed. Your own corrections and pronunciations are applied on the way through, and a per-app profile can change the model and the prompt depending on what you were writing in.
The session captures the field you were in before the recorder even appears, and puts the finished text back into that same field — even if you switched app mid-sentence. Prefer the clipboard? That is a setting, and it puts your old clipboard back afterwards.
Back into the app you started from
Automatic insertion belongs to the direct download, which uses macOS Accessibility. A sandboxed Mac App Store build could only ever reach the clipboard, so the app does not pretend otherwise.
Your words, your keys, your machine
Most dictation apps put themselves between you and a model. This one does not have anywhere to put your audio, because it never has it.
Apple's on-device transcribers, six sizes of WhisperKit running through Core ML, and a local Parakeet model for live streaming. Transcript clean-up can use Apple's on-device Foundation model or a plain offline rules engine. Downloaded once, then no network at all.
Every cloud provider is one you sign up for yourself. The key lives in the Keychain and the app posts your audio straight to that provider's own endpoint. There is no relay, no proxy and no first-party backend anywhere in the codebase — and you can check, because the codebase is public.
No account, no login, no subscription. History lives on your device, and syncs through your own private iCloud only if you switch it on. Diagnostics and anonymous analytics are opt-in and never carry transcripts, audio or keys.
Two honest caveats. Apple's older Speech Recognition can fall back to Apple's servers when on-device recognition is not available on your Mac, and turning on iCloud sync or a cloud voice does send data off the device by design. The full privacy policy spells out each one.
Pick your engine
Speech recognition is moving fast, so the app treats the model as a setting rather than a personality. Switch provider per session, compare two on the same audio, and see what each one cost you.
19 streaming models. Words appear as you speak them.
Thirteen providers plus two families of models you download and run yourself. Drop in a voice memo, a meeting recording or an old interview.
116 built-in voices, plus whatever your provider account lists. Two of these providers are wired up on iPhone.
Optional clean-up pass that fixes punctuation and filler, or follows a prompt you wrote. Cloud ones route through your OpenRouter key.
Counts are of the shipped model catalogue, Sources/SpeakCore/ModelCatalog.swift. Availability varies by platform and by macOS version — iPhone shows a smaller set.
iPhone & iPad
The version you use walking down the street. Start it without unlocking, stop it without looking, and find the words when you sit down.
For people who script things
Three ways in, one vocabulary. Every surface talks to the running app, so your keys, profiles and model choices stay in one place — and no automation surface ever sees a credential.
# A voice memo, straight to the clipboard
$ speak transcribe memo.m4a | pbcopy
# Dictate a commit message
$ speak listen
# ... talk ...
$ git commit -m "$(speak stop)"
# The same verbs, for a coding agent
$ claude mcp add justspeaktoit -- speak mcpThe local socket the CLI and MCP server use is off until you turn it on in Settings → General → Automation. Anything running as you could otherwise start your microphone.
speak CLI
transcribe, listen, stop, history, status. Every command takes --json and prints one versioned envelope for success and failure alike, with distinct exit codes, so scripts can branch properly.
speak mcp exposes start_dictation, stop_dictation, transcribe_file and get_history over stdio to Claude Code and any other MCP client. Calls are idempotent, so a retrying agent cannot open a second microphone.
justspeaktoit:// with start, stop, toggle and dictate, a per-capture destination, language, model and maximum duration, and x-callback-url so Drafts, Shortcuts and iOS 17 automations get the text back. Any app can open the scheme, so the microphone only starts with the app in front of you.
Inside the app
Every one of these settings exists because dictation kept getting the same thing wrong.
Open source
Swift and SwiftUI, MIT licensed, built in the open by one person who uses it every day.
The whole app — Mac, iOS, the shared core and the automation kit. Clone it, run make run, and it is your build.
A provider you already pay for and cannot use here? Open an issue with the template and it goes on the list.
mitchelmore.uk — notes on building this and other things, from the person who makes it.
The other app. A calmer clipboard manager for the Mac, from the same workshop.
Free, open source, and yours the moment it finishes downloading.