MAC DICTATION APP
Roundhand
A menu-bar dictation app for Apple silicon Macs. Hold a key, talk, let go, and the text lands at your cursor, written for the app that has focus.
60-Second Demo
Overview
Speech recognition gives you what you said, not what you meant. A raw transcript has filler words, restarted sentences and missing punctuation, and fixing that by hand erases most of the time you saved over typing.
Roundhand does two things between your voice and the text field. First, recognition: NVIDIA's Parakeet model turns audio into a transcript on your Mac, on the Neural Engine, through CoreML. Speech is recognised on your Mac. Dictation audio is never uploaded or saved to disk.
Second, cleanup: the transcript is cleaned and shaped for the destination app. That step is where I measured offline rules against two cloud LLMs, and the write-up covers what I found.
Apple silicon Mac on macOS 14 or later
Free download, no account needed
Dictation and offline cleanup are free, with no time or word limit
Windows version in development
Key Features
A short message in Slack, a greeting and a sign-off on their own lines in Mail, a list in Notes, one line with no trailing full stop in a terminal.
Parakeet TDT 0.6B v3 runs on the Neural Engine through CoreML. After the one-time model download, the default mode runs on your Mac, apart from update checks.
Deterministic rules drop fillers, keep the version you meant when you restart a sentence, and fix capitals and punctuation. No network, no cost per use, and the same input always gives the same output.
Cloud cleanup is optional. A word-count gate skips the model for transcripts under 6 words by default, because in my sample about a third of the calls returned what the rules had already produced.
A narrow rule turns "or twenty one" into "or-21", so ticket keys survive dictation into a commit message. A bare number in ordinary prose is left alone.
Roundhand will not read or type into password fields.
Frequently Asked Questions
Related Projects
Notes & Credits
Source
Roundhand is closed source and the repository is private. The app is a free download, and the main design decision is written up in the rules vs LLM post.
Author: Avinash Sangle
Website: roundhand.dev
Year: 2026
Built With
- • Swift: the app and its cleanup pipeline
- • NVIDIA Parakeet TDT 0.6B v3: speech recognition
- • CoreML and the Neural Engine: on-device inference
- • Built with Claude Code: AI-assisted development
Try Roundhand for a Day
Hold a key, say a sentence with an "um" and a "scratch that" in it, and see what lands.