A dictation app that actually runs offline

What it does

OpenWhispr is a cross-platform desktop app for voice-to-text dictation, meeting transcription, and AI-assisted note-taking. Hit a global hotkey, speak, and text appears at your cursor. It runs on macOS, Windows, and Linux via Electron, and can operate entirely offline using local speech recognition models.

The interesting bit

The project doesn’t force a cloud subscription on you. Core features — transcription, speaker diarization, semantic search, even AI reasoning — work with local models (Whisper, NVIDIA Parakeet via sherpa-onnx, llama.cpp) or cloud providers you bring yourself. The meeting transcription feature auto-detects Zoom, Teams, and FaceTime calls, then applies on-device speaker fingerprinting so it can label who’s talking without sending audio anywhere.

Key highlights

• Local-first architecture: whisper.cpp and Parakeet run on-device; no telemetry or data collection claimed
• BYOK cloud support: GPT-5, Claude, Gemini, Groq, or local LLMs for the AI agent
• Meeting transcription with live speaker diarization and Google Calendar integration
• Notes with semantic search, folders, and cloud sync (OpenWhispr Cloud backed by Neon Postgres)
• Public API and MCP server for programmatic access and AI assistant integration
• Built with React 19, TypeScript, Tailwind CSS v4, Electron 41, and better-sqlite3

Caveats

• Requires Node.js 24+ to build from source; the README doesn’t specify hardware requirements for local model inference (GPU memory, etc.)
• “GPT-5” is listed as a supported model, which is either forward-looking or a typo — the README doesn’t clarify
• The distinction between fully-offline operation and “OpenWhispr Cloud” features (sync, etc.) could be clearer

Verdict

Worth a look if you dictate regularly and are tired of SaaS lock-in or sending your meetings to someone else’s server. Less compelling if you just need occasional transcription and don’t want to manage local models.