Lexicon based Model for Continuous Speech Recognition in Indian English

Abstract

India has over 150 different languages and over 2000 dialects, according to linguists. Indian English is the second most widely spoken language in India. It does, however, have a lot of varied pronunciations, some separate syntax, and a lot of lexical variance. Their native language and educational backgrounds have a big impact on their pronunciation. A continuous speech recognition model based on the Kaldi toolbox is provided in this research. In Kaldi, finite state transducers are used in the training and decoding algorithms. Decoding takes place on the HCLG graph, where L stands for lexicons, which are phonetic representations of each word and its variations. LDA, MLLT, and SAT were retrieved from the speech corpus to develop the model MFCC and associated transformations. Further comparisons of monophone and triphone models utilizing the n-gram language model's WER and error in words and sentences revealed a significant improvement in the model'saccuracy.Inthetri3model,the WER was5.60,compared to 13.79 in the monophone type.