Meeting Dominance Estimation
To explore how ICSI’s work on speaker diarization can be extended to describe higher-level characteristics of an interaction that are important to humans, we carried out a set of experiments to determine who was the most dominant speaker in a recorded conversation, at least as measured by how much they spoke. We ran our Meeting Dominance Estimation experiments on 4.5 hours of data from the AMI Meeting Corpus, applying speaker diarization techniques to cluster utterances by speaker, then measuring the total time spent speaking for each. We then compared automated dominance estimation based on speaking time with the judgments of human annotators on the dominance ranking of the people in the meeting, and investigated how inter-annotator agreement (or differences) among the humans — especially about the ranking among the less-dominant participants — affected the evaluation of the automatic system.
Our meeting-dominance research also explored systems-level questions, examining differences in performance between using a single microphone vs. one for each speaker and evaluating the practical performance trade-offs between the execution speed of the system, the amount of manual intervention required to make the system work better, and the distance of the audio sensor from each of the participants.
The Meeting Dominance Estimation Demo compares three methods for estimating speaking time.