Quadrilaterals based phoneme segmentation technique for low resource spoken languages

Phoneme segmentation is an important problem that deals with identifying the boundaries of phonetic units. It is a well known problem that has been addressed by researchers for more than 5 decades. The majority of the work reported belonged to English language and languages with enough data sets are available. The popular methods like Hidden Markov Models and Deep Neural Networks require a huge amount of data since they depend on rigorous training process. Even though, these approaches are successful there are difficulties in the context of under-resourced languages. Single scan segmentation algorithms are pertinent in which segmentation can be done locally by inspecting the shape of a speech signal. These methods are efficient in a way that boundary points can be identified in a single inspection. In this paper, a single scan method that is built on geometrical quadrilaterals is proposed which works by understanding the patterns of a speech signal. The algorithm internally uses the geometrical nature of waveform trajectories where the input speech signal is treated as a sequence of structural components. The segmentation process was tested on English spoken digits of Indian accent and Telugu sentences.