Continuous Hindi Speech Recognition Model Based on Kaldi ASR Toolkit

In this paper, continuous Hindi speech recognition model using Kaldi toolkit is presented. For recognition, MFCC and PLP features are extracted from 1000 phonetically balanced Hindi sentence from AMUAV corpus. Acoustic modeling was performed using GMM-HMM and decoding is performed on so called HCLG which is construted from Weight Finite State Transducers (WFSTs). Performance of both monophone and triphone model using N-gram language model is reported which is computed in term of word error rate (WER). A significant reduction in word error rate (WER) was observed using the triphone model. Further, it was found that MFCC feature provide higher recognition accuracy than PLP feature. Goal is to show the performance of Hindi language using present state-of-the-art (Kaldi) system.