Comment by dymk

I think the special sauce of this isn't the STT/TSS (although it does appear to be quite good) - it's the integration of a real-time reasoning model to actually do something with the raw text of the conversation, and whatever they're doing to hide tool call latency if the model needs to search for something.