April 6, 2026
Spice levels: extreme
Show HN: Ghost Pepper – 100% local hold-to-talk speech-to-text for macOS
Privacy purists cheer, speed snobs clap back, and Linux crashes the party
TLDR: Ghost Pepper is a free, on‑device voice‑to‑text app for Mac that pastes your words without sending anything online. Commenters love the privacy but debate accuracy and speed—some push Parakeet, Linux tools, or built‑in options—turning it into a local‑only hero facing tough love from performance purists.
Ghost Pepper drops a “hold-to-talk” voice typing app that runs completely on your Mac, and the crowd instantly splits into camps. The privacy posse is hyped by the no cloud, no data leaving your machine promise, cheering that this free, local tool might make VC-funded dictation apps sweat. The spicy name? Fans say it fits—this thing hides in your menu bar like a ghost and still brings heat.
But the accuracy wars ignite fast. One commenter swears “Parakeet beats Whisper” for speed and precision if your language is supported, while another Linux fan parachutes in with hyprwhspr, bragging about Cohere Transcribe (requires a graphics card) and throwing shade: larger models “shouldn’t need a cleanup sidekick.” That’s a dig at Ghost Pepper’s second model, which tidies your text by removing “um, uh” and self-corrections. Meanwhile, comparison shoppers ask how it stacks against Hex and long-timer Handy, and the convenience crowd waves in yap dictate for folks who don’t want multi‑gigabyte downloads.
The vibe? Privacy-first folks are thrilled, speed hawks want benchmarks, and Linux diehards flex. Jokes fly about “holding Control to control your data,” and one wag quips the app is “so ghostly it doesn’t even show in the dock.” It’s a classic internet showdown: free local hero vs. the accuracy absolutists—and everyone’s got a hot take.
Key Points
- •Ghost Pepper is a free, 100% local hold-to-talk speech-to-text app for macOS 14+ on Apple Silicon (M1+).
- •It uses WhisperKit (small.en, ~466 MB) for transcription and Qwen 2.5 (1.5B + 3B, ~3 GB) for text cleanup, with models served by Hugging Face and cached locally.
- •The app functions as a menu bar utility that pastes transcribed text, includes customizable cleanup prompts, and can launch at login.
- •No cloud APIs are used and no transcriptions are written to disk; debug logs are in memory only.
- •It requires Microphone and Accessibility permissions and supports enterprise pre-approval via MDM PPPC with provided bundle and team IDs; MIT licensed.