Ramdc: Room-Aware Multi-Device Clustering for Large Scale Teleconferencing

Abstract

In large-scale teleconferencing, many devices can share the same physical room even though each one joins as an independent participant. The resulting loudspeaker-microphone couplings can cause cross-device acoustic echo. To mitigate this issue, we introduce Room-Aware Multi-Device Clustering (RAMDC), an unsupervised initialization step at the start of a teleconference. During initialization, every device simultaneously plays a short chirp, and the aggregate responses are transformed into energy decay curves, embedded by a lightweight Conformer encoder, and clustered with DBSCAN to infer room membership. The resulting cluster labels let each device mute streams from co-located devices which can remove cross-device feedback paths. Evaluated on the GTU-RIR dataset, RAMDC outperforms alternatives in the clustering accuracy, and remains robust even in unseen rooms. Show more