Speaking While Listening
A survey and curated list of full-duplex spoken dialogue systems, the models and architectures that let a voice assistant listen and speak at the same time instead of taking strict conversational turns.
This project audits systems that claim full-duplex capability, the ability to listen while speaking rather than only after the other party stops, and finds the term covers substantially different designs. It organizes the space with three frameworks: an architectural hierarchy describing where the duplex decision is made, from external controllers down to shared latent representations; an interaction ontology describing what kind of interaction the system must handle; and a five-state decision machine, idle, listen, speak, wait and dual, describing what the system is doing moment to moment.
Alongside the survey text, the repository collects public full-duplex models, datasets and benchmarks for researchers building or evaluating their own spoken dialogue systems, with visual summaries of the timeline, hierarchy and cross-system audit.
The paper is accepted to EMNLP 2026's main conference.