frog

Frog is an open-source, fully on-device desktop companion for macOS that provides universal dictation, automated meeting notes with speaker diarization, and interactive voice memory without streaming audio or text to the cloud.

For the Non-Technical Reader

Think of Frog as an interactive desktop assistant—reminiscent of a digital pet—that acts as a privacy-focused alternative to tools like Wispr Flow or Granola. Holding a hotkey lets you speak directly into any active window, whether you are drafting an email or writing a long prompt for a coding agent. Beyond simple dictation, Frog can listen into meetings to transcribe who said what, generate post-call summaries, and remember contextual facts you mention out loud, responding back in a custom voice when queried.

For the Technical Reader

Frog integrates several open-source and on-device machine learning components into a single macOS application:

• Speech-to-Text & Diarization: Uses Apple's native SpeechAnalyzer framework or optional NVIDIA Parakeet TDT v3 converted to Core ML via FluidAudio, combined with FluidAudio speaker diarization models.
• Local Inference Engine: Powered by llama.cpp executing Google's Gemma models dynamically selected based on system RAM allocation (ranging from Qwen2.5 1.5B on lower memory machines up to Gemma 26B-A4B on 40GB+ configurations).
• Summarization & TTS: Leverages on-device Apple Intelligence models for meeting synthesis alongside native AVSpeechSynthesizer API hooks for text-to-speech feedback.

Why It Matters

As voice dictation and meeting intelligence tools become central to daily workflows, standard solutions rely heavily on cloud APIs, raising strict data privacy and latency concerns. Frog demonstrates that local Apple Silicon hardware can manage multi-modal voice stacks—encompassing real-time STT, speaker diarization, and local LLM reasoning—entirely on-device with zero recurring API costs and absolute data sovereignty.

The "Voice AI Space Lab" Idea

The Hands-Free Terminal Co-Pilot: Pair Frog's low-latency local dictation with local developer tooling like Ollama or Claude Code. By assigning custom voice shortcuts to trigger context-aware shell execution, developers can dictate complex terminal pipelines or request real-time code refactoring vocally while keeping their IDE strictly offline.

Check out the full open-source project on GitHub.