How to Turn Voice into Code and Notes Without Clouds and Prying Eyes
Imagine: you're on a call, ideas are flying fast, and writing them down manually means losing the flow. Or another situation: you need to quickly draft some documentation, but your fingers are already tired from typing. Sound familiar? Usually in moments like these, we turn to paid services that send your audio data to their servers, analyze it, and possibly train their models on it.
Recently I came across OpenWhispr. It's an open-source project that tries to solve the dictation and transcription problem for regular users. The main feature here is privacy: everything can run locally on your hardware, whether it's a MacBook with an M-chip or a Linux machine with an NVIDIA GPU.
What this tool can do
OpenWhispr is not just a wrapper around a speech recognition model. It's a full-featured desktop application on Electron that integrates into your workflow.
Dictation right at the cursor
The simplest scenario: you press a hotkey, dictate text, and it immediately appears in the active window. Whether it's VS Code, Slack, or a browser. There's even a separate feature for on-the-fly translation — speak in one language, and text in another gets inserted.
Smart meetings and speaker recognition
The project can automatically detect calls in Zoom, Teams, or FaceTime. But the coolest thing here is local diarization. The program remembers voice "fingerprints" and labels who said what, without sending the recording to the cloud. If you often re-listen to meeting recordings to remember who promised to fix that bug by Wednesday, this will save a ton of time.
AI agents at your fingertips
Inside are agents you can control with voice commands. You can connect heavy hitters like GPT-4 or Claude via API, or use local models through llama.cpp. Press a button, ask "write a brief summary of this text," and the agent delivers the result.
Technical internals
The developers put together a pretty solid stack. The foundation is React 19 and Tailwind CSS v4 for the interface, with Electron 41 under the hood.
For voice magic, they use:
- whisper. cpp for high-performance C++ recognition.
- NVIDIA Parakeet and sherpa-onnx for those who prefer GPU speed.
- better-sqlite3 for storing notes and transcriptions.
Interestingly, the project supports Metal acceleration on macOS, CUDA on NVIDIA, and even Vulkan for AMD and Intel GPUs. This means local transcription won't drag on forever.
How to run it
If you don't want to deal with building from source, they have ready-made builds for all platforms. But we're developers here, we want to get our hands on the code. To run from source, you'll need Node.js 24+.
git clone https://github.com/OpenWhispr/openwhispr.git
cd openwhispr
npm install
npm run dev
By the way, the project has its own MCP server (Model Context Protocol). This lets you connect your notes and transcriptions directly to your favorite AI assistant as external context.
Who will find this useful
I see several categories of people who will appreciate OpenWhispr:
- Developers who keep a lot of logs or notes and want to do it faster.
- Team leads who need to capture meeting outcomes without involving third-party spy services.
- Paranoids (in a good sense) who value self-hosted solutions and data privacy.
The project looks very much alive: almost 5,000 stars on GitHub and active commits. Of course, Electron will eat up a fair amount of RAM, and local models will put a load on the CPU, but that's a fair price for keeping your conversations only on your disk.
If you've been looking for a free and open alternative to proprietary dictation services, OpenWhispr is definitely worth spending an evening testing in real-world conditions.
Related projects