Transcribe

Parakeet (local, Intel GPU with CPU fallback) is the speech-to-text for every skill here. Nothing leaves the machine.

Run it

Easiest, from any shell or any LLM tool: transcribe FILE [--srt] [--out-dir DIR] [--verify] (~/.local/bin/transcribe, which runs the script below in the venv at ~/.local/share/transcribe-venv).

Engines, in order (auto moves on to the next when one fails, and prints why):

1. parakeet-gpu: Parakeet TDT 0.6B v3 on the Intel Arc GPU through parakeet.cpp + Vulkan (~/.local/share/parakeet-cpp/: bin/parakeet-cli, lib/, model/tdt-0.6b-v3-f16.gguf; source and build in ~/Work/tries/parakeet-cpp/). 1 min in ~1.5 s, 3 min in ~2.6 s, 30 min in ~26 s.
2. parakeet-cpu: the same binary and model on the CPU (3 min in ~18 s). Same output, so a GPU-only problem is covered without changing anything else.
3. parakeet-onnx: Parakeet through onnx-asr, a separate runtime, in case parakeet.cpp itself breaks.
4. whisper: faster-whisper (medium.en if downloaded, else small.en). A different model with looser timing, last resort.

Every engine runs as its own process (engines 3 and 4 in the venv), so callers can use any python3, and a crash in one engine can't take the caller down.

Load and unload: each call loads the model (about 1 s), runs, and exits, which releases the GPU. Nothing stays resident.