Introduction

After months of typing nine to twelve hours a day, Arun Brahma’s wrists began to ache by mid-afternoon. The voice-to-text tools he tried all shared two drawbacks: monthly subscription fees and audio uploaded to remote servers. Rather than accept either trade-off, he spent several weekends—and eventually a few months—building his own solution. The result is Purr, a free, open-source menu bar application for macOS that handles dictation entirely on the user’s computer.

What Happened

Purr works with a simple shortcut: hold the Right Option key, speak, and release. The words appear in whichever text field currently has focus. No account is required, and once the models download, the app functions without an internet connection.

Beyond basic dictation, Purr includes Smart Typing, which streams text as you speak. Voice Edit lets you select existing text and request changes out loud, such as asking the app to shorten a specific sentence. Meetings mode captures both your microphone and your Mac’s system audio, labels different speakers, and saves the results as Markdown notes with an optional summary.

Under the hood, Purr is written in Swift and SwiftUI. Speech recognition relies on NVIDIA’s Parakeet models through FluidAudio and OpenAI’s Whisper through WhisperKit, both running via Core ML on Apple Silicon’s Neural Engine. A small CTC model rescores transcripts against a custom vocabulary list, while speaker diarization uses a Core ML version of pyannote and SpeexDSP handles echo cancellation. Summaries and Voice Edit run on Apple’s on-device Foundation Models, or Gemma 4 E2B through llama.cpp on Macs without Apple Intelligence.

Development brought a few unexpected audio problems. Without headphones, sound from the speakers leaked back into the microphone and appeared in transcripts. A simple volume cut removed the echo but also cut out quiet speech, so the developer built a system that measures how much echo was removed and preserves quieter words. Voice Edit initially rewrote text the user never asked to change; the fix was to have the model output only specific word replacements, which Purr applies itself. Finally, the first key press caused a two-second delay because macOS started audio services cold; Purr now initializes those services in the background at launch.

Why This Matters

Most cloud-based transcription services store audio on someone else’s server, which creates privacy concerns for anyone handling sensitive notes, emails, or meeting recordings. Purr demonstrates that small open models can deliver a complete dictation workflow locally, without subscriptions or account creation. For writers, developers, and professionals who spend long hours at the keyboard, it offers a hands-free alternative that keeps data on the device where it was created.

It also shows that the hard work in voice software often happens around the model, not inside it—filtering echo, managing audio sessions, and validating edits before they reach the screen.

Key Takeaways

  • Purr is a free, open-source macOS menu bar app for offline voice dictation.
  • It runs on Apple Silicon Macs with macOS 14 or later and requires no account.
  • Features include real-time Smart Typing, voice-activated text editing, and meeting transcription with speaker labels.
  • The stack combines Parakeet, Whisper, Core ML, and Apple’s on-device Foundation Models or Gemma 4 E2B via llama.cpp.
  • All processing stays on the Mac, so audio never leaves the device after the initial model download.

Conclusion

If typing has started to take a physical toll, or if you simply prefer keeping your voice data private, Purr is worth trying. It can be downloaded from purr.arunbrahma.com or built from source on GitHub under an MIT license. The project is a practical example of how small AI models can power everyday productivity tools without relying on the cloud.