I’ve hosted one on a raspberry pi and it took at most a second to process and act on commands. Basic speech to text doesn’t require massive models and has become much less compute intensive in the past decade.