Whisper load balancing / preference lists?

[tonydamage]

Hi,
not sure if this is correct place:
in my local setup I run 2 wyoming-whisper instances

• tiny-8 on Raspberry Pi 4, which is fast and good enough for simple voice commands for Home Assistant
• medium on desktop PC with RTX 2070

The medium whisper is not always on, and what I would like to see is behaviour like this:

• If medium whisper is available (online) use that for speech detection
• Otherwise fallback to tiny-8

It is possible to somehow configure / create wyoming instance which will have preferences? I can contribute code / help developing it, but I would need pointers to starting point and maybe some hints (since stt is stream based, it is harder to decide when switch processing node).

I also have same configuration for piper, but piper proxy is easier for understanding boundaries of request (eg. when I can safely use different instance for generation).

[synesthesiam]

The expected event flow for speech-to-text can be found here: https://github.com/rhasspy/wyoming?tab=readme-ov-file#speech-to-text