Language Identification Based on the Variations in Intonation Using Multi-classifier Systems

Abstract

In this article we make use of the characteristics of tonal languages and machine learning methodologies to understand the patterns in them. Instead of analyzing the absolute pitch or frequency, we analyze how one tone transitions to another in speech. Features (namely, zero crossing count, short time energy, minimum formant frequency, maximum formant frequency) are extracted using the tonal transitions over segments of audio signals. We have developed a multi-classifier system using four classifiers, namely maximum likelihood estimate (MLE), minimum distance classifier (MDC), k-nearest neighbor (kNN) classifier and fuzzy k-NN classifier to automatically identify tonal languages from audio signals. Initially, each individual classifier is trained with existing known data represented by the extracted features. The trained classifier is then used for language identification. Results obtained from these classifiers are combined to generate the final output. Experiments are conducted using three different tonal languages, namely, Chinese, Thai and Vietnamese. The output reveals that the developed multi-classifier model is able to produce promising results. The extracted features produced better results in comparison to usually used frequency value (as a feature). Ensemble of classifiers is a better tool than using individual classifiers.