On latency

There is a perceptual threshold below which a voice interaction stops feeling like a phone call and starts feeling like a tool. We have measured the threshold across hundreds of test sessions. It is approximately 100ms.

Below 100ms turn latency, the rep stops thinking about the model. They stop preparing the cadence of their next sentence with the model in mind. They start listening, the way they would in a real call.

Above 100ms, even by 50ms, the rep adapts. They lean forward. They wait. The buyer becomes a tool to be operated; the practice becomes a different practice — the practice of operating a tool, not the practice of selling.

We have rewritten the speech pipeline three times to fight latency. The current pipeline supports a flash voice profile that routes audio end-to-end without a transcript bottleneck. The latency you pay above 80ms is not algorithmic, it is networking. It is the round-trip from the headset, to a webRTC relay, to the model, back. We pre-warm the route. We pre-allocate the connection. We push DNS into the page load before the rep clicks Start.

If you are evaluating sales-training products, the question to ask is not whether the buyer voice sounds real. It is whether the model interrupts the rep at the right time. Every interruption is a latency budget under 100ms or above it. The buyer who waits a full second to react is not a real buyer. The buyer who interrupts is.