Show HN: Privacy-First Voice-to-Text for macOS (github.com)

WhisperClip is a new open-source macOS app that delivers privacy-first, local voice-to-text using Apple's WhisperKit and MLX (Apple Silicon) optimizations. It runs all speech recognition and LLM-based text processing on-device (no network calls except optional model downloads from Hugging Face), supports multiple Whisper models (216MB–955MB) for speed/accuracy trade-offs, auto language detection, real-time waveform visualization, and customizable prompts for grammar fixing, translation, and other workflows. Local LLM options include Gemma, Llama, Qwen, Mistral, Phi, and DeepSeek variants for on-device grammar correction and text improvement, with global hotkey (Option+Space), auto-copy/auto-paste, menu bar integration, and a polished dark UI. Significance: WhisperClip provides a practical alternative to cloud ASR/LLM services for privacy-conscious users and enterprises—reducing latency, eliminating telemetry, and keeping audio/text on-device while still offering advanced post-processing via local LLMs. Technical implications include Apple-specific dependencies (macOS 14+, Apple Silicon optimizations), ~20GB free disk recommended for model storage, sandboxed execution with encrypted model storage, and an MIT license for commercial/derivative use. Developers can build from source, add models, or contribute improvements; the repo includes build and notarization scripts and a clear setup guide for permissions and model management.