A Canonicalization of Distinctive Phonetic Features to Improve Arabic Speech Recognition

Abstract

The robustness of speech classification and recognition systems can be improved by the adoption of language distinctive phonetic feature (DPF) elements that can increase the effective characterization of a speech signal. This paper presents the results of applying Hidden Markov Models (HMMs) that perform Arabic phoneme recognition in conjunction with the inclusion and classification of their DPF element classes. The research focuses on classifying Modern Standard Arabic (MSA) phonemes within isolated words without a language context. HMM-based phoneme recognition is tested using 8, 16, and 32 HMM Gaussian mixture models. The monophone configuration is designed with consideration of 2-gram language model to evaluate the inherent performance of the system. The overall correct rates for classifying DPF element classes for the three versions of HMM systems are 83.29% 88.96%, and 92.70% for 8, 16, and 32 HMM Gaussian mixture model systems, respectively.