Nanyang Technological University Speaker Diarization in Meetings Domain

The purpose of this study is to develop robust techniques for speaker segmentation and clustering with focus on meetings domain. The techniques examined can however be applied to any other domains. Traditional techniques for speaker diarization developed for telephone conversations or broadcast news are based on a single channel input, which is notably different from meetings domain which can have multiple channel inputs. These techniques when adapted to meetings domain however perform poorer than expected since they do not exploit direction of arrival information, which is available with the presence of multiple microphones. Moreover, many of these techniques are coupled with tunable parameters that need to be manually adjusted for each data set accordingly. In this thesis, our focus is on robust and accurate speaker diarization techniques in meetings. Our aim is to improve the segmentation and clustering performance in diverse conditions while keeping the number of manually tuned parameters to minimal. A comparative study of various distance metrics is first carried out for the purpose of finding the most appropriate metric to use for speaker segmentation and clustering. Our proposed metric is shown to outperform other popular metrics as demonstrated in a speaker verification task. It can be seen as an extension of cross likelihood ratio (CLR) by exploiting the second-order statistic. As an advantage, the computation of this metric is much faster than that of Bayesian Information Criterion (BIC). This work also proposes novel cluster validation methods to determine the optimal number of speakers. In our methods, we define the seperation between sets of intracluster distances and inter-cluster distances as the clustering quality and two metrics are proposed to measure this seperation. For each hypothesized number of speakers, the clustering quality is computed and the maximum value is corresponding to the optimal