A classifier combination approach for Farsi accents recognition

Accent classification technologies directly influence the performance of automatic speech recognition (ASR) systems. In this paper, we evaluate three accent classification approaches: Phone Recognition followed by Language Modeling (PRLM) as a phonotactic approach; accent modeling using Gaussian Mixture Models (GMM) then selecting the most similar model using Maximum Likelihood algorithm that is categorized in acoustic approaches a novel classifier combination method which is proposed to improve the performance of accent classification for several regional accents. In the proposed approach, we use an ensemble method in which each base classifier is a binary classifier that separates an accent from another one. We use the majority vote algorithm to combine the base classifiers. Results for five accents selected from FARSDAT speech database show that the proposed ensemble method outperforms PRLM and GMM-based approaches in the case of Farsi regional accent classifications.