Hacker News Reader

HomeAssistant already does

Hmm yeah but it needs 30 seconds to give an answer. And that's with the LLM running on a GPU with HBM2 memory.

It's not "Gen"erative, but is still a Transformer, and is kinda OK: https://github.com/openai/whisper

(I'm not too impressed with the error rate, but OTOH Alexa and Siri mishear me fairly often as well, so Whisper may be "good enough").

yes. that was the original design actually. and it worked. latency was a bit much though.

There are tools to make your own.

its feasible. if you want to try yourself you can start with the april-asr package.