Press Fn, speak naturally, release. Your words land at the cursor -- polished, in any language, in any app on your Mac.
Native macOS · SwiftUI + Accessibility APIs · Bring your own AI
Quick Start · Providers · Architecture · Contributing
AsterTypeless turns your voice into text and drops it right where your cursor is. No switching apps, no copy-paste, no friction.
You speak --> ASR transcribes --> LLM polishes --> Text appears at cursor
"帮我回复说周三可以" --> "帮我回复说周三可以" --> "周三可以的，到时见。" --> [typed into Slack]
The entire pipeline runs in under 2 seconds with a self-hosted model.
macOS has built-in dictation. It works. But it doesn't clean up your speech, doesn't translate on the fly, doesn't let you swap in your own models, and doesn't work with self-hosted AI. AsterTypeless does all of that.
- Fn-triggered -- hold Fn to talk, release to finish. Also supports tap-to-toggle, double-tap for hands-free, and keyboard shortcuts.
- Writes back to any app -- uses Accessibility APIs to insert text at the cursor. Falls back to clipboard paste when AX isn't available.
- Bring your own provider -- OpenAI, Qwen, Groq, Deepgram, Azure, or any self-hosted server with an OpenAI-compatible API (vLLM, Ollama, etc.).
- Four modes -- Dictate (clean up speech), Rewrite (polish selected text), Translate (into any language), Ask (answer questions from context).
- Minimal floating bar -- a tiny waveform capsule appears while recording. No clutter, no chrome.
- Dark mode native -- respects system appearance out of the box.
Requirements: macOS 14+, Xcode 16+, XcodeGen
git clone https://github.com/AsterZephyr/AsterTypeless.git
cd AsterTypeless
xcodegen generate
./Scripts/deploy.sh # builds, signs, deploys to /Applications, launchesFirst launch opens a permission wizard for Microphone, Accessibility, and Input Monitoring.
Open Settings (gear icon in main window or menu bar) and fill in your endpoint:
| Field | Example |
|---|---|
| Base URL | http://your-server:8000/v1 |
| API Key | sk-... or not-needed for self-hosted |
| Model | gpt-4o-mini / qwen-plus / /data0/models/Qwen3-1.7B |
Click Save -- takes effect immediately, no restart needed. Hit Test Connection to verify.
You can also configure via plist if you prefer:
cp Config/Runtime.sample.plist Config/Runtime.local.plist
# edit with your values| Provider | Chat (LLM) | Transcription (STT) | Streaming | Notes |
|---|---|---|---|---|
| OpenAI | gpt-4o-mini | gpt-4o-transcribe | -- | Standard API |
| Qwen | qwen-plus, qwen-turbo | -- | -- | Via DashScope, OpenAI-compatible |
| Groq | llama-3.3-70b | whisper-large-v3-turbo | -- | Fast inference |
| Deepgram | -- | nova-2 | WebSocket | Real-time streaming |
| Azure OpenAI | Any deployment | Any deployment | -- | Custom endpoint format |
| Self-hosted | Any | Any | -- | vLLM, Ollama, LocalAI, etc. |
Providers with an OpenAI-compatible /v1/chat/completions or /v1/audio/transcriptions endpoint work out of the box.
┌──────────────────────────────────────────────────┐
Fn press ─────────► │ AudioInputMonitor │
│ Records PCM via AVAudioEngine │
Fn release ───────► │ Exports WAV │
└──────────────┬───────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ StreamingTranscriptEngine │
│ POST /v1/audio/transcriptions ──► ASR server │
└──────────────┬───────────────────────────────────┘
│ transcript
▼
┌──────────────────────────────────────────────────┐
│ QuickActionEngine │
│ POST /v1/chat/completions ──► LLM server │
│ (strips <think> tags for reasoning models) │
└──────────────┬───────────────────────────────────┘
│ polished text
▼
┌──────────────────────────────────────────────────┐
│ AccessibilityBridge │
│ AXUIElement.setValue ──► target app cursor │
│ Fallback: clipboard paste via Cmd+V │
└──────────────────────────────────────────────────┘
Sources/AsterTypeless/
├── App/ State machine, bootstrap, pipeline orchestration
├── Features/
│ ├── Home/ Main window (sidebar + capture hero)
│ ├── FloatingBar/ Minimal waveform overlay (NSPanel)
│ ├── Settings/ Provider config, permissions, appearance
│ ├── MenuBar/ Menu bar quick actions
│ └── Onboarding/ First-launch permission wizard
├── Models/ Domain types (sessions, quick bar state, settings)
├── Services/
│ ├── OpenAIClient HTTP: chat completions + audio transcription
│ ├── DeepgramStreamingClient WebSocket: real-time STT
│ ├── ProviderRegistry Multi-provider config + persistence
│ ├── AudioInputMonitor AVAudioEngine capture, PCM buffer, WAV export
│ ├── AccessibilityBridge AX read/write + clipboard fallback
│ ├── HotkeyBridge Fn key tap/hold/double-tap via CGEventTap
│ └── ... Transcript store, config service, shortcuts
└── Support/ Theme tokens, window chrome
# Regenerate Xcode project after adding files
xcodegen generate
# Build
xcodebuild -scheme AsterTypeless build
# Deploy (preserves macOS permissions across rebuilds)
./Scripts/deploy.sh
# Check pipeline logs after a dictation
cat /tmp/aster_pipeline.logThe deploy script uses rsync to update /Applications/AsterTypeless.app without breaking the code signature, so macOS permissions (Accessibility, Input Monitoring) persist across rebuilds.
- Deepgram real-time streaming during recording (live partial transcripts)
- Expanded cross-app compatibility testing (VS Code, Notion, Arc, Slack)
- SwiftData migration for local storage
- Archive, notarization, and DMG distribution
- Configurable Fn behavior (sensitivity, double-tap timeout)
PRs welcome. If you're adding a new provider, it only needs to implement the OpenAI-compatible chat or transcription endpoint format.