← all posts site ↗

my server ended up with one stt and one tts

ran the full diagnostics first, disk, ram, logs, containers, and it was mostly fine, swap active, moderate memory pressure, a couple of dead portal services in the journal. then the actual question: how many speech engines does my server run, and why. the answer was embarrassing. ten plus. fluid, phonon, several parakeet copies, whisper, four gguf benchmark models, plus a text to speech model that never shipped anything.

the benchmark directories were the bulk of it. stt ladder with all its gguf weights, tts benchmark, phonon benchmark, parakeet redux experiments, nine and a half gigabytes of compare and forget. the huggingface cache held another seven and a half gigs of models, only one of which is actually used. a couple of whole apps were dead weight too, a location history stack with its database, and the first two attempts at a console app while a third tauri version is the real one.

cut it down to exactly one stt and one tts on the server, the fallback pair the voice pipeline uses when the desktop machine is slow. everything else deleted. twenty three gigabytes back, four containers gone, and a crash looping service that had restarted over a thousand times disabled. the funniest discovery: the desktop pc keeps its own primary stt and tts, the server only carries backups, so most of what got deleted was running on nothing.

the lesson is that benchmark relics were becoming a second operating system. every model family that lost a comparison should have been deleted the day it lost. one stt and one tts per machine is a policy now, not a preference.