Publisher : Third International Conference on Emerging Research in Electronics, Computer Science & Technology (ICERECT 2018), PES College of Engineering, Mandya
Year : 2018, 2019
Abstract : This paper discusses recognition of digits in Kannada language using an open-source speech recognition tool, Kaldi. The system considers small digit corpora with numbers ranging from 0 to 9 and 4480 samples with a set of fourteen speakers. The monophone and triphone models of the corpora for Kannada language are investigated, and a significant decrease in the word error rate is observed while using the triphone modelling using Kaldi toolkit. The feature extraction is done with Mel-frequency Cepstral coefficients (MFCC). The word error rate (WER) is achieved for this corpus and compared with that achieved the HTK toolkit. A WER of 9% and 6%, respectively, is achieved with monophone modelling and triphone modelling using Kaldi toolkit, and with the same corpus, a WER of 10% is applied in HTK toolkit. A preliminary research of the speech recognition using Kaldi toolkit is reported in this paper. © 2019, Springer Nature Singapore Pte Ltd.