Effective background data selection for SVM-based speaker recognition with unseen test environments: more is not always better

摘要

This study focuses on formulating a procedure to select effective negative examples for the development of improved Support Vector Machine (SVM)-based speaker recognition. Selection of a background dataset, or a collection of negative examples, is the crucial step for building an effective decision surface between a target speaker and the non-target speakers. Previous studies heuristically fixed the number of negative examples used based on available development data for performance evaluation; nevertheless, in real applications this does not guarantee sustained performance for unseen data, as will be shown. In the proposed model selection framework, a novel ranking method is first exploited to rank order the negative examples for selecting a set of background datasets with various population sizes. Next, an error estimation and model-selection criterion are proposed and employed to select the most suitable target model among the model candidates. The experimental validation, conducted on the NIST SRE-2008 and SRE-2010 data, demonstrates that the proposed background data selection slightly but consistently outperforms the fixed-size background data selection, and achieves a relative improvement of +6 % over the non-selection background framework in terms of minDCF.