GMM adaptation based online speaker segmentation for spoken document retrieval

摘要

This paper proposes an online speaker segmentation approach based on Gaussian Mixture Model (GMM) adaptation for spoken document retrieval. In the conventional approach using the Bayesian Information Criterion (BIC), two single Gaussian models are respectively constructed for two divided speech streams in an analysis window, and the dissimilarity between the two models is estimated according to the BIC principle. This approach has been widely applied to speaker segmentation. However, its performance may deteriorate when speakers change frequently, since the single Gaussian model hardly represent the speaker's explicit characteristics for short speech data. To overcome this limitation, we propose an approach to use adapted GMMs instead of single Gaussian models. The method proposed herein constructs a local UBM for speech in an analysis window and adapts the local UBM to each of two divided speech streams in the same window. Upon the two adapted GMMs obtained from the adaptation, the likelihood of the respective speech stream is estimated and change of speaker is determined according to our criterion based on local maxima of BIC. On speaker segmentation experiments based on HUB4, a well-known broadcast news corpus, the proposed method exhibited superior performance compared to the conventional approaches.