Various Speaker Recognition Techniques Using a Special Nonlinear Metric

摘要

In this paper we compare voice recognition techniques using a special nonlinear metric. We describe text-dependent and text-independent speaker recognition methods based on a mel-cepstral analysis in the feature extraction stage and a supervised classification. The Hausdorff-based metric proposed by us is able to measure the distance between different sized speech feature vectors resulted from the mel-cepstral featuring process. Then, a minimum mean distance classifier uses this new distance to identify the speakers.