LocalBolo

LocalBolo is a menu bar app that turns speech into text in any app: Mail, Slack, your code editor, a browser tab. Speech recognition runs on your Mac's Neural Engine and optional cleanup runs on its GPU, so your voice never leaves your Mac. There's no account and no server, and no internet connection is needed once the license is activated and the models are downloaded.

Features

• Hold fn to dictate, anywhere. Release the key and the transcript is pasted at your cursor. fn keeps working as a normal modifier: a quick tap, or fn plus another key, doesn't dictate.
• Shows where your words will land. A small pill at the bottom of the screen grows into a live waveform next to the icon of the app that will receive the text.
• On-device and fast. NVIDIA's Parakeet transcribes 38 seconds of speech in about 0.3 s on an M1.
• Cleans up as you talk (optional). Filler sounds are dropped and a small language model applies your self-corrections: "let's meet at 9 p.m., sorry, 10 p.m." becomes "Let's meet at 10 p.m."
• Clipboard-safe. Text is pasted through the clipboard, which is then restored.
• Choose your models. Parakeet by default, or OpenAI Whisper in three sizes.
• Dictate in English, Chinese, Hindi or Japanese. Choose a language or automatic detection in Settings. Non-English dictation uses Whisper Turbo, entirely on your Mac.
• Polish your writing (optional). Fix grammar and punctuation and add paragraphs or lists with a local multilingual model. Keep conservative English cleanup when you prefer minimal edits.

Screenshots

First-run setup	Speech models	Transcript cleanup

System requirements

Chip	Apple Silicon (M1 or newer). Intel Macs aren't supported: Parakeet and MLX only run on Apple Silicon, and the app is built for arm64 only.
macOS	15 Sequoia or later
Memory	8 GB or more. Usage depends on the selected speech and cleanup models.
Disk	About 60 MB for the app, plus models: Parakeet ~450 MB, and optionally the cleanup model ~840 MB
Keyboard	One whose fn / 🌐 key macOS can see, such as a MacBook keyboard or an Apple Magic Keyboard. Many third-party keyboards handle Fn internally and never report it.
Permissions	Microphone (to record while fn is held) and Accessibility (to see fn from any app and paste with ⌘V)
Network	To activate the license key and download models the first time; afterwards only for daily update checks and a license re-check about every two weeks
Language	English, Mandarin Chinese, Hindi and Japanese; automatic language detection with Whisper Turbo

Getting started

Signed releases will be published on the Releases page. Until then, build the app from source.

Build from source

You need Xcode 26 or later with the Metal Toolchain, which MLX uses to compile its GPU kernels:

xcodebuild -downloadComponent MetalToolchain
git clone https://github.com/Thinking-Sound-Lab/localbolo.git
open localbolo/apps/mac/LocalBolo.xcodeproj

Choose the LocalBolo Dev scheme and run. On the first build Xcode resolves the Swift packages and asks you to Trust & Enable the CudaBuild plugin from mlx-swift. It does nothing on a Mac, but Xcode asks about every package plugin.

First launch

LocalBolo lives in the menu bar and has no Dock icon. A setup window walks you through:

1. Microphone access.
2. Accessibility access.
3. Downloading the Parakeet speech model (about 470 MB).
4. Optionally setting System Settings › Keyboard › "Press 🌐 key to" to Do Nothing, so macOS doesn't also open the emoji picker when you release fn.

The first launch after installing or updating takes about 45 seconds while the speech model is optimized for your Mac's Neural Engine. After that the app is ready in about 2 seconds.

Transcript cleanup

Turn it on in Settings › Cleanup or from the menu bar. Existing settings keep the Conservative (English) style, which downloads Qwen 2.5 1.5B (about 880 MB) and runs on the GPU with MLX:

• Rules first. "Um", "uh" and accidentally doubled words ("the the", "can you can you") are removed by plain rules, which take no time and can't change what you said.
• Only where needed. The model only reads sentences where you may have corrected yourself ("sorry", "I mean", "scratch that"). Every other sentence is pasted exactly as spoken, so a long dictation is cleaned up as quickly as a short one.
• Fast. An edit takes about 0.2 s on an M1. The model reads its instructions once, when it loads, and checks your dictation as a draft instead of writing it out again word by word.
• Guarded. An edit is used only if every change is one a correction explains: words you took back may go, and nothing new may appear. Otherwise (for example if the model answers a question you dictated) that sentence is pasted as you said it.

Choose Polished to fix grammar, spelling and punctuation, remove fillers and apply spoken corrections throughout the dictation. It uses Qwen 3 1.7B (about 990 MB) and can add paragraphs and plain-text lists. The prompt asks it to preserve meaning, names, numbers and each language in mixed-language speech, without translating or answering questions. Checks reject obvious content changes; failed or unfinished edits fall back to the original passage. These checks cannot prove that every edit preserves meaning.

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗