Conversational Infill Hides Reasoner Latency in Voice Agents Using a Small Talker Model

A small real-time talker model generates contextually grounded filler responses immediately while a slower reasoner model runs in parallel, then fluently integrates the reasoner's streamed output mid-response. Trained on a 290,571-example synthetic dataset across six domains, the approach is validated across seven small models, decoupling latency from capability in voice agents.