Most speech demos show you a transcript. The interesting engineering is one layer down: the socket. This is a streaming recognition endpoint viewed from the protocol side — interim hypotheses arriving and being revised before anything is committed, endpointing deciding where an utterance stopped, per-frame latency measured as percentiles rather than an average, and a decode queue that applies backpressure when a client streams faster than realtime. Switch it to live microphone and the VAD, adaptive noise floor and endpointer run on your actual audio.