Unsupervised speaker indexing of discussions using anchor models

Abstract

AbstractWe present unsupervised speaker indexing, combined with automatic speech recognition (ASR) for speech archives, such as discussions. Our proposed indexing method is based on anchor models, by which we define a feature vector based on the similarity with speakers of a large‐scale speech database. We introduce dimensional normalization and reduction on the vectors to improve discriminant ability. These vectors are then clustered and initial speaker labels are obtained. Using the initial labels, speaker models are constructed for respective clusters and the speakers are finally indexed with the speaker models. We perform ASR using the results of this indexing. We achieved a speaker indexing accuracy of 97% and a significant improvement in the ASR for real discussion data. © 2005 Wiley Periodicals, Inc. Syst Comp Jpn, 36(9): 25–33, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20215