Whisper Turbo vs Moonshine Voice: local dictation setup

This walkthrough compares Moonshine Voice and Whisper Turbo for dictating prompts into Claude Code and Telegram conversations with OpenClaw. The presenter tests speech recognition on his own machine, then builds a Python script that types the transcript into the focused window.

Moonshine starts in an isolated Python environment with a tiny English model downloaded on first use. The presenter tries small and medium models after transcription errors, but finds the results inadequate for his often-mumbled speech. These informal tests reflect his setup and speaking style, rather than a controlled accuracy benchmark.

The Windows Whisper setup uses FFmpeg, PyTorch with CUDA support for his Nvidia GPU, and SoundDevice for microphone access. He describes the GPU as helpful rather than mandatory. His script handles recording and silence detection, stopping after two seconds of quiet; the initial version limits recordings to 30 seconds. Whisper processes the recording after speech ends, while Moonshine provides live transcription.

The final local AI dictation script runs in the background and uses Ctrl+Shift+Space. Global hotkeys require an administrator terminal in this setup, and microphone noise calibration needs attention. The presenter prefers Whisper's accuracy despite its heavier installation. One dictation trial takes about 25 seconds versus roughly 40 to 45 seconds for typing. He reports that transcription needs no cloud service or API key.

The transcription runs locally; Telegram messaging and the connected coding or chat agents are separate services and are not established as offline.