Summary
Neil Zeghidour, co-founder and CEO of Gradio (the Paris-based startup behind the Moshi full-duplex speech model and the Hibiki real-time translation system), traces the full arc of voice agent technology and argues that today’s systems are fundamentally constrained by half-duplex turn-taking — the walkie-talkie problem of the talk’s title.
The presentation moves from Siri’s 2011 constrained NLU pipeline through OpenAI’s original conversational voice mode to current “third generation” voice agents, demonstrating a live drive-thru ordering bot built on Gradio’s models. The core technical argument is precise: full-duplex conversation requires the model to generate and receive audio simultaneously, but raw audio at 24kHz produces 72,000 time steps for an eight-word sentence — roughly 10,000 times longer than the equivalent text sequence — making direct LLM processing 100 million times more computationally expensive due to quadratic attention complexity. Neural codecs that compress audio into discrete tokens are the enabling technology.
Zeghidour illustrates why latency alone cannot solve the walkie-talkie problem: a live demo shows a half-duplex model repeatedly interrupting natural backchanneling sounds (“mhm,” “yeah”) because it cannot distinguish acknowledgment from turn-taking. In human conversation, both parties speak simultaneously up to 20% of the time. The talk positions full-duplex modeling — which Gradio released as Moshi in 2024 — as the necessary architectural direction for voice agents that feel genuinely natural.
📺 Source: AI Engineer · Published September 15, 2026
🏷️ Format: Deep Dive







