Audio-based unsupervised segmentation of multiparty dialogue

In this paper, we explore a novel way to leverage audio information for unsupervised segmentation of multiparty dialogue. Our system which segments directly on patterns derived from audio sources is evaluated with previous work that segments on lexical patterns found in transcripts. We examine the effectiveness of both systems on recovering a two-layer structure of meeting dialogue. We demonstrate that the audio-based system performs significantly better than the word-based system on this task. In particular, it effectively recover segments of off-topic discussion. Results are encouraging as the audio information used in the system can be obtained in near real time and with absence of manual and ASR transcripts. It is particularly desirable when a system has to be operated online, or in unfamiliar domains and languages.