Ghost Pepper: Local WhisperKit Transcription and LLM Refinement

The Pitch

Ghost Pepper is an open-source macOS utility developed by Matt Hartman that provides 100% local dictation via a hold-to-talk hotkey (GitHub). It uses WhisperKit for initial transcription and a secondary local LLM pass to clean up and format the resulting text, positioning itself as a privacy-centric alternative to cloud-heavy GPT-5 voice inputs.

Under the Hood

The tool utilizes WhisperKit (Argmax) for local Apple Silicon inference, ensuring that no audio data leaves the machine (GitHub Codebase). It relies on a ~1GB transformer model footprint, which has drawn some criticism from the developer community for being "bulky" compared to more efficient 2026 models like Parakeet TDT (600MB) used in competing apps (BGR News; Hacker News).

While the logic is sound, Ghost Pepper is fighting an uphill battle against macOS Tahoe. Recent benchmarks show that Tahoe’s native Liquid Glass dictation overlays and on-device AI are currently 55% faster than standard third-party Whisper implementations (UsedBy Dossier).

Competition in the local ASR space is currently at a peak. Cohere Transcribe, released in March 2026, currently leads the Open ASR Leaderboard with a 5.42% Word Error Rate (WER), making Ghost Pepper’s accuracy claims harder to justify without more transparent benchmarking (Reddit/Hugging Face).

We don't know yet which specific local LLM is being used for the post-transcription cleanup phase. The documentation does not specify if it is a quantized Llama-4-mini or a custom Qwen implementation, nor do we have data on the battery drain compared to the native Tahoe Dictation API.

Marcus's Take

Ghost Pepper is a well-engineered project that arrives in a hyper-saturated market where the OS vendor has already won. While the "local LLM cleanup" pass is a nice touch for formatting, it doesn't justify the overhead of a 1GB model when macOS Tahoe handles this natively and with better power efficiency. It is essentially a member of a growing "support group" of independent apps trying to outrun Apple's vertical integration (Hacker News). Skip this for production use; the native APIs or Cohere’s recent release are objectively superior for 2026 workflows.

Ship clean code,
Marcus.