Discriminative Approach for Building ASR of Kashmiri Language

Speech recognition is the hottest form of AI research these days, but many under resources languages still haven't traversed yet. This book deals with one of undewr resource language: Kashmiri language. Authors describe the development as well as implementation of CMU Sphinx-3 based speech recognizer for the Kashmiri language. Recognition of the words has been done by using hidden markov models (HMMs). Dictionary data consists of one (akh) to hundred (hat) Kashmiri digits, most frequent Kashmiri words and sentences from the book "Spoken Kashmiri a Language Course" by Omkar N. Koul. Here, we developed a speaker independent, Kashmiri - Automatic Speech Recognition (K-ASR) system for continuous speech. System is trained using the speech corpora that contains 47 Kashmiri sentences, a total of 2800 words spoken by 12 male and female speakers in noise as well as in noise-free environmental conditions. Maximum Accuracy of 89.78% was achieved, when the K-ASR system was tested in noise free environment.