Dialect classification via discriminative training

Variability in speech due to dialect is a major factor limiting speech system performance for speech recognition, spoken document retrieval, and dialog systems. In this study, we propose a novel discriminative algorithm to improve dialect classification for unsupervised spontaneous speech in Arabic. No transcripts are used for either training or testing, and all data are spontaneous speech. The Gaussian mixture model (GMM) is used as our baseline system for dialect classification. The major motivation is to remove confused/distractive regions from the dialect acoustic space, while emphasizing discriminative/sensitive information. The Kullback-Leibler divergence is used to find the most discriminative GMM mixtures (KLD-GMM), after which the confused acoustic GMM region is removed. The proposed algorithm is evaluated on three dialects of Arabic, with measurable improvement achieved (4.0%), over a generalized maximum likelihood estimation GMM baseline (MLE-GMM) system.