wersper

Private, self-owned Wispr Flow. Hold Right ⌥, speak, release — the transcription is pasted at your cursor. Everything runs locally on your Mac (Whisper large-v3-turbo via MLX on the GPU). No cloud, no accounts, no telemetry.

Run

uv run wersper.py

First run downloads the model (~1.6GB from Hugging Face) and installs deps.

macOS permissions (one-time)

Grant your terminal app both of these in System Settings → Privacy & Security:

• Microphone — to record you
• Accessibility — to paste into other apps (simulated ⌘V)

Self-check

uv run wersper.py --check

Notes

• Esc while recording cancels the take.
• Pop = recording started, Tink = text pasted.
• Your clipboard is restored after pasting (text only — images are lost).
• Hotkey, model, language, and silence threshold are constants at the top of wersper.py. Set LANGUAGE = None for autodetect.