Improved Classification Of Acoustic Features Via Primal Weight Vectors

This paper presents a new variation of the Support Vector Machine (SVM) technique for speaker identification in audio. Primal Weight Vectors (PWV) provide a discriminative framework to distinguish between SVM models. Here we discriminate between Gaussian Mixture Model Super Vector (GSV) models, which represent one of the leading SVM based approaches to speaker identification. The PWV-GSV combination has demonstrated a consistent performance advantage over the state of the art GSV classifier on a variety of conditions.