Audio-Video detection of the active speaker in meetings

Abstract

Meetings are a common activity that provide certain challenges when creating systems that assist them. Such is the case of the Speaker recognition, which can provide useful information for human interaction modeling, or human-robot interaction. Speaker recognition is mostly done using speech, however, certain visual and contextual information can provide additional insights. In this paper we propose a speaker detection framework that integrates audiovisual features with social information, from the meeting context. Visual cue is processed using a Convolutional Neural Network (CNN) that captures the spatio-temporal relationships. We analyse several CNN architectures with both cues: raw pixels (RGB images) and motion (estimated with optical flow). Contextual reasoning is done with an original methodology, based on the gaze of all participants. We evaluate our proposal with a public benchmarks in state-of-art: AMI corpus. We show how the addition of visual and context information improves the performance of the speaker recognition.