Your Voice Agent is Just a Walkie Talkie — Neil Zeghidour, Gradium
Neil Zeghidour from Gradium explains that today's voice assistants work like one-way radios — they can either listen or speak, never both at once.
In normal conversations, people overlap their speech about 20 percent of the time. Zeghidour traces the evolution from 2011's simple voice assistants through open-ended voice conversations without agency, to agents that can execute tasks but suffer from latency. The core problem is that raw audio requires enormous computational power — eight words equal 72,000 processing steps — so models use compressed audio instead. To achieve truly natural voice conversations requires models that handle two audio streams simultaneously, but every improvement in naturalness has so far required less intelligence.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
