Your voice, unheard by others.
Privacy-first voice-to-text for macOS and Windows, built with Rust.
Murmur is a voice dictation tool that transcribes your speech and inserts polished text at your cursor position -- in any app. It supports both local (on-device) and cloud transcription, with optional LLM post-processing to clean up filler words, fix punctuation, and convert Simplified Chinese to Traditional Chinese.
- Push-to-Talk -- Hold a modifier key to speak, release to insert text
- Toggle Mode -- Press once to start recording, press again to stop (with 5-min auto-stop and debounce protection)
- Custom Hotkey -- Single modifier key or combo (e.g. Option+Z, Control+Space) with two-phase recording
- Dual Engine -- Local Whisper (Metal GPU) or Groq cloud API
- Multi-Provider LLM -- Groq (cloud), Ollama (local), or any OpenAI-compatible endpoint for text enhancement
- Fully Offline Mode -- Local Whisper + Ollama for complete privacy (no data leaves your machine)
- LLM Post-Processing -- Clean up filler words, add punctuation, Simplified-to-Traditional Chinese conversion
- Translate Hotkey -- Select text in any app, press a hotkey (default: Option+T) to translate and replace in-place via LLM
- Smart Clipboard -- Auto-pastes when a text field is focused; copies to clipboard only when no text input is detected (e.g. on Desktop)
- App-Aware Style -- Automatically adjusts output tone based on the active app (e.g. formal in Slack, technical in VS Code)
- Personal Dictionary -- Add custom terms to improve transcription accuracy; inline dictionary chips appear in real-time while editing
- Transcription Preview -- Floating preview window with copy button, editable text, character count, and detected app name
- Live Preview -- See partial transcription while you speak (local engine only)
- Mixed-Language Support -- English words in mixed CJK-English speech are preserved as-is (never translated)
- 15 Languages -- Auto-detect or manually select from 15 supported languages
- Cross-Platform -- macOS and Windows support with platform-native hotkey and app detection
- System-wide -- Works in any text field across all apps
- Lightweight -- Tauri-based, ~30-50MB vs 200MB+ Electron apps
- Open Source -- Fully auditable, no telemetry, no tracking
Download the latest release from the Releases page.
| Platform | File | Notes |
|---|---|---|
| macOS (Apple Silicon) | .dmg |
Requires quarantine removal |
| Windows | .exe / .msi |
CPU-only, works on all hardware |
| Windows (NVIDIA GPU) | -cuda.exe / -cuda.msi |
GPU-accelerated via CUDA |
Voice: Hotkey -> Record (cpal) -> Transcribe (Whisper) -> LLM Clean-up (optional) -> Smart Clipboard
Translate: Option+T -> Copy selection -> LLM Translate -> Paste back (replace selection)
Each recording triggers at most 2 API calls (when using Groq): one for Whisper transcription, one for LLM post-processing. Translation triggers 1 API call.
git clone https://github.com/panda850819/murmur-voice.git
cd murmur-voice
pnpm install
pnpm tauri devOn first launch, Murmur will guide you through:
- Granting Microphone and Accessibility permissions
- Choosing a transcription engine (Local or Groq)
- Setting your Push-to-Talk key
If you choose the local engine, the Whisper model (~1.5GB) will download automatically on your first recording.
| Engine | Speed | Quality | Privacy | Setup |
|---|---|---|---|---|
| Local (Whisper) | ~1-3s | Good | Audio stays on device | Model auto-downloads on first use (~1.5GB) |
| Groq API | <1s | Good | Audio sent to Groq servers | Free API key (get one below) |
To switch engines: Settings > Transcription > Engine
- Go to console.groq.com and sign up (Google/GitHub login supported)
- Navigate to API Keys in the left sidebar
- Click Create API Key, give it a name (e.g. "murmur")
- Copy the key (starts with
gsk_) and paste it into Murmur's settings
Groq's free tier includes generous rate limits for personal use. The same API key is used for both Whisper transcription and LLM post-processing.
Choose a provider for AI text enhancement:
| Provider | Speed | Privacy | Setup |
|---|---|---|---|
| Groq | Fast | Cloud | Free API key from console.groq.com |
| Ollama | Varies | Local | Install Ollama, pull a model |
| Custom | Varies | Varies | Any OpenAI-compatible endpoint |
What it does:
- Removes filler words (um, uh, etc.)
- Removes false starts and self-corrections
- Adds proper punctuation (full-width for Chinese, half-width for English)
- Converts Simplified Chinese to Traditional Chinese (Taiwan standard)
- Adds spaces between Chinese and English text
- Formats lists and paragraphs when appropriate
To enable: Settings > AI Processing > LLM Post-Processing
Select text in any app and press Option+T (default) to translate it in-place. The translated text replaces your selection and appears in the preview window.
- Uses the same LLM provider configured in AI Processing (Groq, Ollama, or Custom)
- Target language configurable in Settings > Translation > Target Language
- Translate hotkey customizable in Settings > Translation > Translate Hotkey
- Preview window stays visible until you close it or click Copy
Add frequently used terms (names, jargon, acronyms) to improve transcription accuracy. These are injected into Whisper's initial prompt.
To configure: Settings > Transcription > Dictionary (type a term, press Enter to add)
When enabled, Murmur detects the foreground app and adjusts the LLM output tone:
| App | Style |
|---|---|
| Slack, Discord, LINE, Telegram | Casual |
| VS Code, Terminal, Cursor | Technical |
| Pages, Word, Google Docs | Formal |
| Others | Default (natural) |
To enable: Settings > AI Processing > App-Aware Style
For the best experience with Chinese dictation:
| Setting | Value | Why |
|---|---|---|
| Engine | Groq | Fastest transcription (<1s) |
| Language | Mandarin Chinese | More accurate than Auto for Chinese |
| LLM Post-Processing | On | Cleans up filler words + Traditional Chinese |
| LLM Model | Llama 3.3 70B | Best quality for Chinese text processing |
| App-Aware Style | On | Adapts tone to context |
| Component | Technology | Purpose |
|---|---|---|
| App Framework | Tauri 2 | Lightweight desktop app |
| Audio Capture | cpal | Microphone input -> 16kHz mono |
| Speech-to-Text | whisper-rs / Groq API | Local or cloud transcription |
| LLM Processing | Groq / Ollama / Custom | Text cleanup and formatting |
| Hotkey Detection | CGEventTap / SetWindowsHookEx | Global hotkey listener (modifier or modifier+key combo) |
| Text Insertion | arboard + rdev | Clipboard write + Cmd+V / Ctrl+V simulation |
| App Detection | NSWorkspace / Win32 API | Foreground app detection (per-platform) |
- macOS 12.0+ (Apple Silicon recommended for local Whisper)
- Microphone permission
- Accessibility permission (for global hotkey + text insertion)
- Windows 10+
- Microphone permission
- Groq API key (free, for cloud engine and Groq LLM) or Ollama (for local LLM)
This happens because the app is not signed with an Apple Developer certificate. macOS Gatekeeper quarantines unsigned apps by default. To fix:
- Move Murmur Voice to
/Applications - Open Terminal and run:
xattr -d com.apple.quarantine /Applications/Murmur\ Voice.app - Open the app normally
| Your GPU | Download | Why |
|---|---|---|
| NVIDIA (with CUDA drivers) | -cuda version |
GPU-accelerated transcription, much faster |
| AMD / Intel / integrated | Standard version | CPU transcription, works on all hardware |
| Not sure | Standard version | Always works, just slower for local engine |
Murmur is a free, open-source project. Apple Developer Program costs $99/year. Code signing may be added in the future, but for now the workaround above is required on macOS.
Murmur was born from a security audit of a commercial voice-to-text app that was found to:
- Capture browser URLs and window titles
- Monitor all keystrokes via CGEventTap
- Send application context to remote servers
- Include session recording analytics (Microsoft Clarity)
Murmur does none of this. When using the local engine, your audio never leaves your machine. When using Groq, audio is sent only to Groq's API for transcription -- no other data is collected or transmitted.
If you find Murmur useful, consider supporting the project:
Crypto:
| Network | Address |
|---|---|
| EVM (Ethereum, Base, etc.) | 0x9ae8954201b2fce97b124887e415df02e8e06a8d |
| Solana | Eod4VqvMmmMnY3EinN6Zo5xzt9Wq5S2dFZutob1VBvMf |
MIT