Performance of classifiers on MFCC- based phoneme recognition for language identification

Abstract

The automatic identification of language from voice clips is known as automatic language identification. It is very important for a multi lingual country like India where people use more than a single language while talking making speech recognition challenging. An automatic language identifier can help to invoke the language specific speech recognizers making voice interactive systems more user friendly and simplifying their implementation. Phonemes are unique atomic sounds which are combined to constitute the words of a language. In this paper, the performance of different classifiers is presented for the task of phoneme recognition to aid in automatic language identification as well as speech recognition. We have used Mel Frequency Cepstral Coefficient (MFCC) based features to characterize Bangla Swarabarna phonemes and obtained an accuracy of 98.17% on a database of 3710 utterances by 53 speakers.