Full-Duplex Voice AI Has No Standard Definition. This New Survey Proposes 3 Frameworks to Change That.

The term 'full-duplex' has been liberally applied to over a dozen spoken dialogue systems, yet it describes vastly different capabilities. Existing surveys tend to flatten these differences onto a single axis—cascaded vs. end-to-end or engineered vs. learned—missing distinctions crucial for system builders. To address this, a team of 17 researchers from leading institutions introduces three complementary frameworks. The L0-L3 Architectural Hierarchy specifies where in the pipeline duplex decisions are made, from level 0 (no duplex) to level 3 (joint representation-level modeling). The T×I×R Interaction Ontology defines interactions by three coordinates: temporal relation (e.g., barge-in, overlapping), user intent, and required system response. Finally, the Decision State Machine (IDLE, LISTEN, SPEAK, WAIT, DUAL) models moment-by-moment behavior transitions.

Auditing published systems and benchmarks, the researchers document a striking 'realization gap': many architectures can theoretically operate in full-duplex states, but their observed behavior remains constrained by the interaction patterns present in training and evaluation datasets. Public training data covers only a fraction of the interaction types represented in (largely undisclosed) industrial corpora. The paper identifies the still-unrealized goal of L3 representation-level modeling—where the system jointly models speech production and comprehension at the deepest level—as the key frontier. For builders, these frameworks offer a precise vocabulary to design, compare, and benchmark true full-duplex systems, ultimately pushing voice AI toward interruption-capable, natural conversation.