Murmur

Local voice-to-text for Windows. Hold a key, talk, let go — your words land wherever you're typing.

---

Everything runs on your machine. No cloud, no account, no sending audio anywhere. Murmur sits in your system tray and gives you a global hotkey that turns speech into text in any app — your editor, your browser, a chat window, whatever has focus.

How It Works

graph LR
    A["🎤 Hold hotkey"] --> B["🎙️ Mic capture"]
    B --> C["📡 WebSocket"]
    C --> D["🧠 Transcription engine"]
    D --> E["💬 Partials stream to overlay"]
    E --> F["📋 Release → clipboard + paste"]

Audio flows from your mic through an AudioWorklet, gets sent as 16-bit PCM over a local WebSocket, and hits the transcription engine running on your GPU (or CPU). Partials stream back in real-time so you see words forming as you speak. When you release the key, the final transcription lands in your clipboard and gets pasted automatically.

The overlay is a transparent, always-on-top, click-through window — it shows up when you're recording and gets out of the way when you're not.

• Windows 10/11

• Bun

• Python 3.11+

• uv

• just

• CUDA-capable GPU recommended (driver 525+; CPU is supported but slower)

• Hold-to-talk or toggle mode — bind any key as your global hotkey

• Transparent overlay — live waveform and partial transcription while you speak

• Two startup-selectable engines — Nemotron by default, Whisper for multilingual use

• Auto-paste — transcribed text goes straight to your clipboard and into the active field

• Post-processing — auto-append periods, spaces, or both

• Searchable history — every transcription saved locally in SQLite

• In-app server controls — start, stop, restart, stream logs, all from the settings panel

• External server mode — point Murmur at a remote server if you want

Engines

Murmur ships with two transcription engines. Both run locally; choose the engine when starting the server with MURMUR_ENGINE or server/settings.json.

	Nemotron	Whisper
Model	nvidia/nemotron-speech-streaming-en-0.6b	large-v3-turbo (via faster-whisper)
Best for	English dictation, low latency	Multilingual, accuracy
Streaming	Native streaming architecture	Chunked re-transcription
Extras	—	Hotword boosting
VRAM	~1.5 GB	~3 GB

Nemotron is the default. It's a 0.6B parameter model built for streaming — partials come back fast and the final result is usually identical to the last partial. Start the server with Whisper for non-English languages or when you need hotword support to nail domain-specific terms.

Quick Start

You need Windows 10/11 with Bun, Python 3.11+, uv, and just. A CUDA GPU is recommended but not required.

# Server
cd server
uv sync --extra all    # or: --extra nemotron / --extra whisper
just start

# Start with an explicit engine
pwsh -NoProfile -Command "$env:MURMUR_ENGINE='nemotron'; uv run python -m main"
pwsh -NoProfile -Command "$env:MURMUR_ENGINE='whisper'; uv run python -m main"

# App (separate terminal)
cd app
bun install
bun run dev

The app auto-detects a running server in dev mode. In production, it manages the server lifecycle itself.

Note

If you develop from WSL, run all uv/bun/just commands through PowerShell — not Linux. Running them from WSL replaces Windows binaries with Linux ones and breaks everything. See BUILDING.md.

Build

cd server && uv sync --extra all
cd ../app && bun run package:win

bun run package:win produces a small nsis-web installer stub plus payloads (for example .7z, .yml, .blockmap) in app/release/. End users need internet to install (payload download) and to fetch models on first run.

Root-level helper:

just build

See BUILDING.md for full release and troubleshooting details.

Configuration

App settings (hotkey, audio device, post-processing, auto-paste) are configured through the UI.

Server settings use MURMUR_-prefixed environment variables or server/settings.json. Engine-related settings are read at server startup.

Server environment variables

Variable	Default	Description
MURMUR_HOST	0.0.0.0	Bind host
MURMUR_PORT	51717	Bind port
MURMUR_ENGINE	nemotron	Default engine (nemotron / whisper)
MURMUR_NEMOTRON_MODEL	nvidia/nemotron-speech-streaming-en-0.6b	Nemotron model
MURMUR_NEMOTRON_DEVICE	auto	Device (auto/cuda/cpu)
MURMUR_WHISPER_MODEL	large-v3-turbo	Whisper model
MURMUR_WHISPER_DEVICE	auto	Device (auto/cuda/cpu)
MURMUR_WHISPER_COMPUTE_TYPE	auto	Whisper precision mode
MURMUR_MAX_SESSIONS	10	Concurrent session cap
MURMUR_LOG_LEVEL	INFO	DEBUG/INFO/WARNING/ERROR

Project Structure

app/      Electron desktop app (Svelte 5, TypeScript, Tailwind v4)
server/   Transcription server (FastAPI, faster-whisper, NeMo)
docs/     Protocol spec and technical docs

Protocol

The app and server communicate over a custom WebSocket protocol on port 51717 — binary frames for audio, JSON frames for control and text. Full spec: docs/protocol.md

License

MIT