wanted to know how hard it would be to swap the speech engines in the desktop dock, whisper for stt, piper for tts, instead of the two that ship. the honest answer is that the seam is already http, every consumer, the voice router, home assistant, knows only two urls, one for transcription and one for speech. swap what answers on those ports and nothing upstream notices. that is the whole reason this is a small job instead of a rewrite.
but the count says otherwise. six places assume one engine. the service builder has one launcher per kind, the config has one flat struct holding every knob from every engine, the model list reads one folder per kind, the installer has hardcoded rows and one installer function each, the command layer has over twenty config lookups, and the ui keys on kind not id. that last one rules out the naive add another service path, a second stt service would collide with the first because the ui returns one of them.
the correct design makes the engine a property of the service. add an engine field to the config and replace the builder match with a table that maps engine to executable, arguments, health url and environment. the impact stays small, about six files with a handful of edits and a couple of tests. the result keeps the same five services on the same ports, with the engine selectable per service.
the tts side is already free. the generated config declares two voice families, so switching between them is a dropdown, not a migration. everything else is packaging.