Continuous Speech Recognition System for Maithili

Abstract With the increasing role of speech technology in society, it has become indispensable for researchers to develop systems for low-resource languages that can cater to the needs of the common man residing in remote areas. This paper presents the inaugural attempt to develop an automatic continuous speech recognition system in Maithili, a low-resource language spoken by around 80 million people in India and Nepal. Leveraging a speech corpus from the Linguistic Data Consortium for Indian Languages containing 21412 speech utterances covering 300 speakers, the baseline speech recognition trained using the Kaldi toolkit employs the traditional Hidden Markov Model (HMM) in tandem with the Gaussian Mixture Model(GMM). Feature extraction utilizes Mel Filter Cepstral Coefficient (MFCC) and the ASR system is trained using GMM and Deep Neural Network (DNN) along with HMM. Language model training incorporates n-gram modeling. The system is trained on both Devanagari and Mithilakshar scripts, achieving the best Word Error Rates (WER) of 16.31% and 19.09% respectively.