[NOTE] vibe coded local STT because Handy microphone settings are really bad and and I can't configure it to work better so I made my own with codex agents - basically 100% ai-written

handyclone

• hotkey client that records mic audio, cleans it, and posts it to a FastAPI ASR server using NVIDIA Parakeet; the transcript can be auto-typed into the active window.

requirements

• python 3.12+, cuda-capable gpu for the server model download/load
• microphone access; speakers for feedback beeps
• deps tracked in pyproject.toml + uv.lock

install (uv + lock)

uv venv
source .venv/bin/activate
uv sync  # installs from the lockfile

run the asr server

python asr_server.py

• downloads and loads nvidia/parakeet-tdt-0.6b-v3 to cuda, exposes POST /asr at port 8000.

run the client

python asr_client.py --server http://localhost:8000/asr

• hold ctrl + command (left or right) to record; release to stop
• default flow: bandpass + rms normalize, save temp wav, send to server, print transcript, type it into the focused app
• useful flags: --raw (skip audio cleanup), --no-type, --type-on keyword (only type if keyword present), --delay seconds between keystrokes, --delay-per-word, --debug-keys