Reed Dictation: free, local-only voice dictation for macOS. Hold a shortcut, talk, release.
See the codeFree, private voice dictation for Mac. Hold a shortcut, talk, release: clean text appears at your cursor, in any app. Everything runs on your Mac.
| You say | Reed types |
|---|---|
| "Um, so I think we should, uh, ship it on Friday." | So I think we should ship it on Friday. |
| "Let's meet at seven PM, sorry, five PM." | Let's meet at 5:00 PM. |
| "Call me on Tuesday, I mean Wednesday." | Call me on Wednesday. |
| "The invoice came in at three thousand one hundred and sixty two dollars, due on the fifteenth of March." | The invoice came in at $3,162, due on the 15th of March. |
| "I'm sorry, I can't make it today." | I'm sorry, I can't make it today. |
Real output of Reed 0.3.2 on a Mac with Apple Intelligence, from speech made with the Mac's built-in voice. Fillers go, self-corrections resolve, numbers, amounts and dates are written the way you would type them, and what you meant to say stays. Without Apple Intelligence, a rules-based cleanup does the same kind of tidying.
Comparing options? See Reed next to Wispr Flow, Superwhisper and Apple Dictation.
Requires macOS 14 or later on Apple Silicon (M1 or newer). AI cleanup needs macOS 26 with Apple Intelligence; other Macs use the rules-based cleanup automatically. Reed checks for updates itself.
Dictation has no network path: no audio, no transcripts, no requests. The app's network gate (GateURLProtocol) enforces it, and you can confirm it with any network monitor, such as Little Snitch.
The network is used for two things only: downloading the speech model (from Hugging Face) and checking for updates (Sparkle, from Reed's release bucket). Like any web request, those carry your Mac's IP address, and the update check names the app and Sparkle versions in its user agent. Details: what Reed collects.
While you talk, Reed splits the recording at pauses and processes each piece on the Mac: noise reduction, speech recognition with NVIDIA's Parakeet v3 on the Neural Engine, a deterministic pass for numbers and vocabulary, then cleanup with Apple's on-device Foundation Models. Every model edit is checked word by word against what you said, and refused if it rewrites your words instead of tidying them. On release, the pieces are joined and typed at your cursor.
Which languages does it understand? English, today. Other languages are not supported yet.
Does it work offline? Yes. After the one-time model download, dictation and cleanup need no connection.
Can I use the Fn key? Not alone. Use a modifier hold (like the default ⌃⌥) or a modifier plus a letter.
Does it work in Electron and web apps? Yes. Where an app does not accept typed text through Accessibility, Reed pastes instead and then restores your previous clipboard (unless you copied something new in the meantime).
Do I see text while I talk? Not yet. The menu bar icon shows recording and processing, and the text appears when you let go.
Does it run on Intel Macs? No. The speech model needs Apple Silicon.
Bug reports and ideas are welcome. Report reproducible bugs with the bug template, and bring ideas and questions to Discussions. Reed's maintainers implement accepted changes themselves, so unsolicited pull requests are closed; CONTRIBUTING.md explains how to take part.
To build Reed, run the tests, or find your way around the code, see docs/DEVELOPMENT.md.
If you are an AI agent helping with this repository, read AGENTS.md before making changes and CONTRIBUTING.md before opening issues or pull requests. Project rules for reviews and verification live in CLAUDE.md; they apply to people and to AI-assisted sessions alike, and every pull request uses .github/PULL_REQUEST_TEMPLATE.md.
Reed is licensed under the Apache License 2.0. Copyright 2026 W20 Labs Inc.; see NOTICE.
Reed Dictation: free, local-only voice dictation for macOS. Hold a shortcut, talk, release.
See the codeFree, private voice dictation for Mac. Hold a shortcut, talk, release: clean text appears at your cursor, in any app. Everything runs on your Mac.
| You say | Reed types |
|---|---|
| "Um, so I think we should, uh, ship it on Friday." | So I think we should ship it on Friday. |
| "Let's meet at seven PM, sorry, five PM." | Let's meet at 5:00 PM. |
| "Call me on Tuesday, I mean Wednesday." | Call me on Wednesday. |
| "The invoice came in at three thousand one hundred and sixty two dollars, due on the fifteenth of March." | The invoice came in at $3,162, due on the 15th of March. |
| "I'm sorry, I can't make it today." | I'm sorry, I can't make it today. |
Real output of Reed 0.3.2 on a Mac with Apple Intelligence, from speech made with the Mac's built-in voice. Fillers go, self-corrections resolve, numbers, amounts and dates are written the way you would type them, and what you meant to say stays. Without Apple Intelligence, a rules-based cleanup does the same kind of tidying.
Comparing options? See Reed next to Wispr Flow, Superwhisper and Apple Dictation.
Requires macOS 14 or later on Apple Silicon (M1 or newer). AI cleanup needs macOS 26 with Apple Intelligence; other Macs use the rules-based cleanup automatically. Reed checks for updates itself.
Dictation has no network path: no audio, no transcripts, no requests. The app's network gate (GateURLProtocol) enforces it, and you can confirm it with any network monitor, such as Little Snitch.
The network is used for two things only: downloading the speech model (from Hugging Face) and checking for updates (Sparkle, from Reed's release bucket). Like any web request, those carry your Mac's IP address, and the update check names the app and Sparkle versions in its user agent. Details: what Reed collects.
While you talk, Reed splits the recording at pauses and processes each piece on the Mac: noise reduction, speech recognition with NVIDIA's Parakeet v3 on the Neural Engine, a deterministic pass for numbers and vocabulary, then cleanup with Apple's on-device Foundation Models. Every model edit is checked word by word against what you said, and refused if it rewrites your words instead of tidying them. On release, the pieces are joined and typed at your cursor.
Which languages does it understand? English, today. Other languages are not supported yet.
Does it work offline? Yes. After the one-time model download, dictation and cleanup need no connection.
Can I use the Fn key? Not alone. Use a modifier hold (like the default ⌃⌥) or a modifier plus a letter.
Does it work in Electron and web apps? Yes. Where an app does not accept typed text through Accessibility, Reed pastes instead and then restores your previous clipboard (unless you copied something new in the meantime).
Do I see text while I talk? Not yet. The menu bar icon shows recording and processing, and the text appears when you let go.
Does it run on Intel Macs? No. The speech model needs Apple Silicon.
Bug reports and ideas are welcome. Report reproducible bugs with the bug template, and bring ideas and questions to Discussions. Reed's maintainers implement accepted changes themselves, so unsolicited pull requests are closed; CONTRIBUTING.md explains how to take part.
To build Reed, run the tests, or find your way around the code, see docs/DEVELOPMENT.md.
If you are an AI agent helping with this repository, read AGENTS.md before making changes and CONTRIBUTING.md before opening issues or pull requests. Project rules for reviews and verification live in CLAUDE.md; they apply to people and to AI-assisted sessions alike, and every pull request uses .github/PULL_REQUEST_TEMPLATE.md.
Reed is licensed under the Apache License 2.0. Copyright 2026 W20 Labs Inc.; see NOTICE.
