Mizo Spoken Query System Enhanced with Prosodic Information

We report the development of a Mizo spoken language interface to an online database of commodity prices in various agricultural markets of Mizoram province. The spoken query system is designed to be used over mobile phone networks. It is inspired by earlier such systems developed for languages such as Assamese and Marathi. However, considering the manifold increase in the performance of the current Mizo language spoken query system, we were motivated to report the whole system implementation in detail. The average word error rate of the DNN-HMM speech recognition system in a 5-fold cross validation experiment is 2.2%. The word error rate of the Mizo commodity name recognition system is 2.75 times smaller than that reported for a similar system for Assamese language. The reduction in error rate is attributed to data collection from diverse environmental situations with different noise levels and further processing of the data. The word error rate of the Mizo ASR system increases to 8.5% in field trials conducted with several users.