EUROSPEECH 2003 - INTERSPEECH 2003
This paper addresses the estimation of a speaker GMM through the selection and merging of a set of neighbors models for that speaker. The selection of the neighbors models is based on the likelihood score for the training data on a set of potential neighbor GMM. Once the neighbors models are selected, they are merged to give a model of the speaker, which can also be used as an a priori model for an adaptation phase. Experiments show that merging neighborhood models captures significant information about the speaker but doesn't improve significantly compared to classical UBM-adapted GMM.
Bibliographic reference. Mami, Yassine / Charlet, Delphine (2003): "Speaker modeling from selected neighbors applied to speaker recognition", In EUROSPEECH-2003, 2629-2632.