Jarvis

Jarvis exists because its author got annoyed that a rival raised 81 million dollars for voice dictation, and spent three months of late nights on an open source answer. Hold the Fn key, speak, release, and punctuated text appears in whatever field the cursor sits in, with fillers such as um and like stripped out automatically.

Transcription and cleanup are separate stages, and either can run locally or in the cloud. Local transcription uses Whisper or NVIDIA Parakeet through Sherpa-ONNX, local language models run through Ollama, and the cloud path pairs Deepgram with Gemini when speed matters. The app carries no tracking or telemetry.

The behavior is not fixed at whatever the author shipped.

• Prompt engineering: every behavior is customizable, from how mail is formatted to how dictation is cleaned.
• Text actions: fix grammar, rephrase, turn speech into bullet points or generate text outright.
• Small commands: simple actions already work, such as opening YouTube or setting a five minute timer.
• Hands-free mode: a double tap on Fn toggles continuous recording, and Escape cancels it.
• Model choice: Whisper in tiny, base or small, or NVIDIA's 600 million parameter Parakeet model.
• Ollama support: pulled models are detected automatically and picked from a dropdown.

Signed and notarized builds are published for Apple Silicon and Intel Macs, an iOS version is on TestFlight, and building from source needs Node 18 or later, macOS 10.13 or later and the Xcode command line tools. A Windows version is on the roadmap, and the license is MIT.