Mizo Phone Recognition System

In this work, we present the creation of a speech corpus in an under-resourced language, namely, Mizo, and report the development of a phone recognition system using state-of-the-art acoustic modelling techniques. The speech corpus consists of Mizo passages spoken by 81 native Mizo speakers. A set of 36 phonetic units have been identified for representing the acoustic phonetic characteristics of the speech signal. Over the past few decades, the Hidden Markov Model (HMM) based acoustic modelling technique is most widely used to capture the temporal variations in the speech data. However, the state observation probabilities in the HMM were generated through multivariate Gaussian Mixture Models (GMMs). Additionally, we have explored two new frameworks in developing the Mizo phone recognition system, namely, the Deep Neural Network (DNN) and the Subspace Gaussian Mixture Model (SGMM). Phone Error Rates (PERs) of 18.1% and 17.5% are observed in DNN-HMM and SGMM-HMM based approaches, respectively, without utilizing any language constraint. PER reduced to 13.9% and 15.7%, respectively, after incorporating the language model information.