brain
projects·source-summary

github-wispr-flow

GitHub README for xaid-vfx/wispr-flow. Description: "Wispr-flow for free." Pinned. 3 stars.

Key extractions

  • Mechanism: Whisper via MLX transcribes, raw output pasted instantly. Say "delta, ..." as a prefix to route utterance through local LLM (Ollama / dolphin3) for cleanup before paste.
  • Hotkey: push-to-talk — hold right ⌘ while speaking, release to paste. Also VAD continuous mode.
  • Config: --hotkey-key, --model (tiny.en → large-v3), --llm-model, --prompt (bias Whisper), --threshold (VAD energy).
  • Requirements: macOS Apple Silicon, Ollama, Python 3.11+.
  • Test harness exists: pytest tests/ mentioned.

Complementary to portfolio

Portfolio source describes the why and architecture (MLX vs PyTorch tradeoff, sub-300ms latency, INT4 quantization). This README is the how to run it and interaction model ("delta" prefix as the clever UX). Together they fully describe the product.

Cross-ref

The "delta, ..." prefix as a magic word for LLM-routed rewrite is a distinctive UX choice worth extracting into compress-mechanical-labor or a prompt-pattern playbook later.