Towards Manipuri Tonal Contrast Disambiguation Using Acoustic Features

Abstract

Manipuri is a low resource tonal language of the Tibeto-Burman language family. Preliminary studies confirm that there are two tones in the Manipuri language: Level tone and Falling tone. For such tonal languages, features that characterize the tone distinctly are essential for developing a robust speech recognition systems. The existing tone-based methods have not studied or analyzed Manipuri tones in this context. Therefore, in this work, we carried out an acoustic feature analysis of the Manipuri speech samples. Firstly, we extend the existing ManiTo dataset containing 3000 samples of isolated Manipuri tonal contrast word by including additional 3000 samples. Secondly, the proposed work extracts ten selected features from each utterance present in the given speech samples. These features are further analyzed for their ability to distinguish the above two mentioned tones. The results validate that our selected features can efficiently differentiate the tones in the Manipuri language.