Neural network classifiers for language identification using phonotactic and prosodic features

Abstract

In this paper, we explore phonotactic and prosodic features derived from the speech signal and its transcription for identification of a language. The characteristics of languages represented by phonotactic and prosodic features at the trisyllabic level are used to train feedforward neural network (FFNN) classifiers to discriminate among languages. We demonstrate that these features indeed contain language-specific information. We also show that phonotactic features in terms of broad phonetic categories are sufficient to represent the phonotactic regularities/constraints of languages. The performance of the FFNN classifier based on these features is evaluated for three Indian languages.