Broad Phoneme Classification - A Study

Different studies done on broad phoneme classification reveal that the realization of a language independent speech system is a challenging task. The main aim of this work resides in the realization of language independent broad phoneme classification. An ANN based language dependent Broad Phoneme Classifier (BPC) is developed for English, Malayalam and Telugu languages. The BPC classifies speech signal into six broad phoneme categories such as Silence, Vowels, Nasals, Fricatives, Approximants, and Plosives, using 16 signal-based input features. The classifier is modified with three additional features in order to address the poor performance of approximant classification. These features are fourth formant, ratio of second to first formant and envelope variance measure. A major improvement in classification is achieved when the system is trained using Deep Neural Network Classifier. Since the features used are signal based, it can be easily be adapted for any language. Effectiveness of the proposed features are demonstrated through language independent broad phoneme classification experiments in three languages.