Recurrent neural network based approach to recognize assamese fricatives using experimentally derived acoustic-phonetic features

Fricatives are the major group of speech sounds bearing distinct acoustical and phonetical characteristics and provides a wide range of application possibilities in the field of speech and speaker recognition. Assamese, which is a widely spoken language in the north eastern part of India, has four distinct fricative sounds called /s/, /z/, /x/ and /h /. In this paper, a Recurrent Neural Network (RNN) based algorithm is described to recognize fricative sounds from Assamese speech, where a feature vector is generated from the specific acoustic-phonetic characteristics like centre of gravity (COG), standard deviation (SD), skewness and kurtosis of the fricatives of the language. The experimental results show around 96% success rate which well above in comparison to the state of art of the phoneme recognition strategy.