Voice demos happen in quiet rooms on good networks with speakers the model has heard before. ATRIUM does not get any of those conditions. It answers calls for a service business: cordless phones, warehouse echo, callers who are outside in the wind, and a network that degrades every day around lunchtime.
This entry is a list of the gaps between the lab and the room, and what we did about each one. None of the fixes are clever. All of them were necessary.
Latency is a budget, not a number
People stop believing they are in a conversation somewhere around a second of silence. Our budget is 800 milliseconds from end of speech to start of reply, and every component, transcription, reasoning, synthesis, network, spends part of it. The discipline is refusing to let any single component spend more than its share, even when the “better” model is 300 milliseconds slower.
Accents, interruptions, and memory
The transcription model shipped with excellent benchmark numbers and still struggled with several of the client’s regular callers. We handle this less with model tuning and more with conversation design: the system confirms names, numbers, and addresses by reading them back, every time, because a fifteen-second confirmation is cheaper than a mis-booked job.
Real callers interrupt. The system has to stop talking the moment a caller starts, mid-sentence, without sulking, without losing its place. And it has to remember the whole call: a caller who gave their address two minutes ago and gets asked for it again hangs up, reasonably.
“The lab condition is not the room. Build for the room, and the lab numbers stop mattering.
After twelve months, the metric we watch most closely is not word error rate. It is the proportion of calls that finish without the caller ever asking for a human. That number, more than any benchmark, tells us whether the loop is fast enough and the memory is good enough.