English | 简体中文
Voice-to-prompt input for AI coding
Flowtype is a macOS voice input app built for AI coding workflows.
It helps developers turn spoken, messy, and highly verbal thoughts into clearer prompts for coding agents like Codex, Claude Code, and similar tools.
- Voice is faster than typing — describe ideas at natural speaking speed
- Spoken thoughts flow better — they're more continuous and expressive than typed text
- Raw transcription isn't enough — speech is too conversational for AI coding tools
- Flowtype bridges the gap — structured, coding-oriented text refinement with one keypress
| Action | Result |
|---|---|
Double press Command |
Start voice recording (capsule window appears) |
Single press Command (while recording) |
Stop and output raw spoken text |
Double press Command (while recording) |
Stop and output LLM-polished structured prompt |
- Describe a feature idea hands-free while reviewing code
- Turn rough implementation thoughts into a usable coding prompt
- Quickly draft UI, workflow, and product instructions for AI coding tools
- Brainstorm architecture decisions out loud, then paste the cleaned result
- Record — Double press
Commandto start voice capture (a capsule window appears at the bottom) - Preview — Apple on-device speech recognition shows real-time transcription as you speak
- Transcribe — Audio is sent to ASR providers (parallel routing with scoring) when recording stops
- Refine — LLM cleans up filler words, fixes recognition errors, and structures the prompt (double-press end only)
- Inject — Result is typed directly into your active text field
Sources/flowtype/
├── App/
│ ├── FlowTypeApp.swift # Entry point, accessory-only app
│ └── StatusBarController.swift # Status bar icon & menu
├── Core/
│ ├── AppState.swift # Global state management
│ ├── Configuration.swift # Configuration model
│ ├── ConfigurationStore.swift # UserDefaults persistence with debounce
│ ├── EnvMigration.swift # One-time .env → GUI migration
│ ├── PipelineOrchestrator.swift # End-to-end audio → text pipeline
│ └── AsyncRefiner.swift # Parallel ASR + LLM refinement
├── Services/
│ ├── AudioRecorder.swift # macOS audio capture (segmented)
│ ├── KeyboardInjector.swift # Text insertion via clipboard / HID
│ ├── LLMService.swift # SiliconFlow SSE streaming client
│ └── Speech/
│ ├── SpeechRouter.swift # Multi-provider routing & scoring
│ ├── SpeechProvider.swift # Protocol
│ ├── ASRPostProcessor.swift # Filler stripping, term correction
│ ├── ASRResultScorer.swift # 7-dimension quality scoring
│ ├── AppleSpeechProvider.swift # On-device speech recognition (preview + fallback)
│ ├── TeleSpeechProvider.swift
│ ├── SenseVoiceProvider.swift
│ └── SiliconFlowSpeechProvider.swift
├── Settings/
│ ├── SettingsView.swift # SwiftUI settings panel
│ └── SettingsWindowController.swift # Settings window host
├── UI/
│ └── AudioVisualizer.swift # Recording visual feedback
├── Utilities/
│ ├── AudioFormatConverter.swift # PCM → WAV, normalization, silence trim
│ ├── SegmentMerger.swift # Deduplicated segment merging
│ ├── DotEnv.swift # .env file parser (legacy)
│ └── PermissionHelper.swift # Accessibility permission check & guide
├── Resources/
│ ├── tech_terms.json # Tech term corrections
│ └── filler_words.json # Filler word dictionary
- macOS 14+
- Swift 6.2+
- SiliconFlow API key (for LLM refinement and ASR)
# 1. Clone
git clone <repo-url>
cd Flowtype
# 2. Build
swift build
# 3. Run
swift run FlowTypeOr build the .app bundle:
./scripts/build-app.sh
open build/Flowtype.appAll settings are managed through the Settings GUI (click the status bar icon → Settings, or press Cmd + ,):
| Section | Settings |
|---|---|
| ASR Primary | Provider, Base URL, API Key, Model ID |
| ASR Fallback | Provider, Base URL, API Key, Model ID |
| LLM | Provider, Base URL, API Key, Model ID |
| Trigger Key | Fn / Control / Option / Command |
| ASR Strategy | Parallel (run both, pick best) or Fallback (primary first) |
Settings are persisted to UserDefaults automatically. An existing .env file will be migrated once on first launch, after which the GUI settings take precedence.
The tools/ directory includes an evaluation framework for benchmarking ASR providers:
cd tools
cp .env ../.env # ensure API key is available
python evaluate_asr.py --output-dir eval_output/See tools/eval_data/README.md for dataset details.
MIT