lamitype

Lamitype (formerly HushType) is a speech-to-text application designed for macOS, with the core positioning of "free, private, Traditional Chinese-first, and doesn't hog memory." By default, it uses the 4-bit MLX quantized version of Qwen3-ASR 0.6B to run entirely locally on Apple Silicon. The model weights are about 708 MB, measured RAM peak is about 2.1 GB, and idle cache is limited to 1 GB. The model can be unloaded from the menu to free memory. It supports voice input in English, Chinese, and Japanese, as well as mixed-language sentences. Through OpenCC s2twp, it provides stable Traditional Chinese output, and includes deterministic ITN to convert Chinese numerals to Arabic numerals. Main features include: press and hold Right Option for voice input; local Live Caption (free, offline); optional Live Translated Caption (connects to OpenAI gpt-realtime-translate, supporting 14 languages); on-device text translation via Apple Translation Framework; and Text Polish in-place proofreading using Apple Foundation Models (requires macOS 26+ and Apple Intelligence). Cloud dictation is an opt-in feature, supporting OpenAI (default gpt-4o-mini-transcribe) and Gemini (default gemini-3.5-flash-lite). It uses the user's own API key to connect directly to the provider over HTTPS, with no relay server in between. After each app launch, consent is requested before the first cloud request, and it includes a daily spending alert and same-day lock guardrail. When switching cloud engines, the local model is not loaded at all; the engine selection is preserved across restarts. Two installation options are provided: download the DMG (fully standalone version, with OpenCC and all dependencies included, no Homebrew or terminal commands needed) or compile from source. iOS support is experimental: the Mac serves as the transcription server, and the iPhone sends audio to the Mac via WiFi or Tailscale; this requires additional installation of Python 3 and related packages. Regarding privacy: local mode does not store recordings, has no telemetry, and can work completely offline; cloud mode has no relay server in the middle, and keys are stored in the local Application Support directory with 0600 file permissions. The Proofreading Lab feature temporarily stores proofreading decisions from local dictation as plain text, deletes them on normal exit, and clears them at next launch after abnormal termination; the folder is not indexed by Spotlight.