voxline

Hold a key. Speak. Get polished writing.

A native macOS dictation app that turns your voice into clean, written text — anywhere on your Mac. Speech runs on-device via Whisper. Cleanup runs through your own LLM API key, so you control the model, the cost, and the data path.

macOS 14.0+ · Apple Silicon · Bring your own API key

     

Note

voxline is in active development. Expect rough edges. The transcription pipeline is solid; the surface around it is still maturing.

What it does

Most dictation tools dump a raw transcript with um, uh, half-finished sentences, and zero punctuation. voxline does the second step you actually want:

1. Speech → text locally on your Mac via Whisper.
2. Text → polished writing through a frontier LLM you choose (Claude or GPT).

Result: hold a hotkey, say what you mean — even messily — and watch clean prose appear in whatever field you're typing into.

Features

• Hold to talk — press your hotkey, speak, release. Text appears in the focused field.
• Works in any text field — browser, email, IDE, terminal, Slack, Notes, Cursor, ChatGPT, anything that accepts a paste.
• On-device transcription — Whisper runs locally on Apple Neural Engine. Audio never leaves your Mac.
• AI cleanup, not raw dump — fillers, false starts, and rambling are smoothed out. Punctuation and capitalization are added automatically.
• Transform selected text by voice — highlight text anywhere, hold your hotkey, and say how to change it: "make this a bullet list", "make this cleaner", "make this shorter". voxline rewrites the selection in place and leaves it as a normal ⌘Z-undoable edit. No selection? It just dictates, as usual.
• Context-aware per-app formatting — voxline detects the frontmost app and tunes the output for it: terse Slack messages, structured email replies, code-comment style in your IDE, search-box one-liners. Ships with sensible defaults for 28 common apps out of the box.
• Dictation history — the last 25 cleaned dictations live in a menu-bar submenu; click any row to copy it back to the clipboard.
• Bring your own LLM key — Anthropic or OpenAI, your account, your model, your costs. Keys live in macOS Keychain.
• Menu-bar native — no Dock icon, no clutter. Configurable hotkey, mic, model, and provider.
• Privacy-aware feedback — clipboard is restored after paste; the system mic indicator turns off the moment you let go.

How it works

┌──────────────── ON YOUR MAC ─────────────────┐
│  🎙 audio  →  Whisper (ANE)  →  transcript   │
│                                              │
│  Context block, assembled at press time:     │
│  • per-app mode prompt                       │
│  • focused-field AX (role, surroundings)     │
│  • custom vocabulary (canonical spellings)   │
└─────────────────────┬────────────────────────┘
                      │  transcript + context
                      ▼
┌────────── CLAUDE or GPT (your key) ──────────┐
│  strip fillers · fix self-corrections ·      │
│  match register · snap to vocab              │
└─────────────────────┬────────────────────────┘
                      │  polished writing
                      ▼
┌──────────────── ON YOUR MAC ─────────────────┐
│  paste into focused field                    │
│  (clipboard restored; AX/typing fallback)    │
└──────────────────────────────────────────────┘

Speech models (on-device)

Model	Size	Best for
Whisper large-v3 turbo (default)	~1.5 GB	Highest accuracy, multilingual
Whisper small.en	~466 MB	Lightweight, English-only, fastest first run

Models download on first use via WhisperKit and are cached locally. Switching models in Settings triggers an on-demand download — no app reinstall.

Cleanup providers (cloud, your account)

Provider	Default model	Where to get a key
Anthropic	claude-haiku-4-5	https://console.anthropic.com/settings/keys
OpenAI	gpt-4.1-nano	https://platform.openai.com/api-keys

Why cloud cleanup instead of a local model? Because the gap between a frontier LLM and what fits on a laptop is still enormous for prose quality. voxline's bet: trust on-device for the audio (which is sensitive), and let you pick best-in-class for the cleanup (which only sees a transcript). You decide which provider.

Requirements

macOS	14 (Sonoma) or later
Mac	Apple Silicon — M1, M2, M3, M4, or any variant. Intel Macs are not supported.
RAM	8 GB minimum, 16 GB recommended (the default large-v3-turbo model is happier with headroom)
Disk	~2 GB free for speech models (large-v3-turbo ~1.5 GB, small.en ~466 MB). Models cache inside the app container.
Network	Required on first launch to download the Whisper model, and at runtime for AI cleanup. Pure transcription works offline once the model is cached.
Microphone	Any input device macOS recognizes (built-in mic is fine).

Apple Silicon is non-negotiable: voxline runs Whisper on the Apple Neural Engine via WhisperKit, and there is no ANE on Intel Macs.

Getting started

Download a signed build from Releases — the DMG is notarized by Apple, and the app checks for updates automatically via Sparkle.

Or build from source:

git clone https://github.com/tfredricks/voxline.git
cd voxline
open voxline.xcodeproj

Build and run from Xcode (⌘R), or use ./scripts/build-local.sh to build Release and install straight to /Applications.

On first launch:

1. Grant Microphone and Accessibility when prompted (the app will guide you).
2. Pick your hotkey, mic, and Whisper model in the Settings window (⌘,). The default hotkey is Right Cmd + Right Option — change it if you'd rather use something else.
3. Drop in an Anthropic or OpenAI API key in the Cleanup (AI) section.
4. Hold the hotkey anywhere on your Mac and start talking.

Permissions

Permission	Why
Microphone	Capture your voice while the hotkey is held. Audio never leaves your Mac.
Accessibility	Detect the global hotkey and paste into the focused field.

Input Monitoring is not required — Accessibility alone is enough for the global hotkey. macOS may still surface an Input Monitoring entry for voxline; you can leave it off. The Debug pane shows its status for diagnostics only.

Privacy

What stays local:

• 🎙 Audio capture — held in memory only, never written to disk, dropped as soon as the transcript exists.
• 🧠 Speech-to-text — runs entirely on Apple Neural Engine via WhisperKit. No audio is sent anywhere.
• 🔑 API keys — stored in macOS Keychain. Not logged, not synced, not visible to other apps.

What goes to your LLM provider:

• ✍️ The transcript only — voxline sends a small text request to Anthropic or OpenAI for cleanup. Whatever provider's privacy policy applies (use enterprise tier or org-level keys if that matters to you).
• voxline has no telemetry, no analytics, and no first-party server. The only network traffic is to whichever LLM provider you choose and the Hugging Face model download on first use.

Good to know

• Sandboxed app — voxline runs inside the macOS app sandbox, so model files live in the container, not your home folder.
• Settings live in one place — single-page Settings window with a status strip up top showing what's wired up. Switching providers keeps both keys around for fast toggling.
• Resilient hotkey — when permissions are revoked or restored, voxline reconciles automatically without a relaunch.
• Open source — read the code, audit the data path, file an issue, send a PR.

Acknowledgments

Built on the shoulders of:

• WhisperKit — on-device Whisper inference for Apple Silicon
• Whisper — the original model from OpenAI
• swift-transformers — model hub and inference utilities
• The macOS dictation tools that paved the way (Whispr Flow, Superwhisper, Ghost Pepper, and others) — voxline borrows the hold-to-talk UX they all converged on.

Releasing

Maintainer notes — the steps to cut a tagged release.

Versioning. voxline uses SemVer for the marketing version (CFBundleShortVersionString — e.g. 0.2.1). Bump it manually when cutting a release:

• Patch (0.2.0 → 0.2.1) — bugfixes only
• Minor (0.2.0 → 0.3.0) — new features, backwards-compatible
• Major (0.x → 1.0.0) — first stable release, or breaking changes after that

The build number (CFBundleVersion) is the commit count on main and stamps itself at build time — never edit it by hand.

Release checklist.

1. Update MARKETING_VERSION in voxline.xcodeproj/project.pbxproj (one line).
2. In CHANGELOG.md, move [Unreleased] items under a new [X.Y.Z] - YYYY-MM-DD heading.
3. Commit: chore: release vX.Y.Z.
4. Tag and push:

   git tag vX.Y.Z
   git push origin main --tags

5. Create a GitHub Release from the tag; paste the CHANGELOG entry as the body.

That's the whole flow. Between releases, MARKETING_VERSION stays put — every dev build reports the last released version with a higher commit-count build number.

Contributing

Bug reports, feature ideas, and pull requests are welcome. See CONTRIBUTING.md for build instructions, testing, the DCO sign-off requirement, and the Code of Conduct.

License

voxline is licensed under the Apache License, Version 2.0.

See NOTICE and THIRD-PARTY-NOTICES.md for required attributions. The "voxline" name and logo are reserved — see TRADEMARK.md.