Improving Large Vocabulary Urdu Speech Recognition System using Deep Neural Networks

Development of Large Vocabulary Continuous Speech Recognition (LVCSR) system is a cumbersome task, especially for low resource languages. Urdu is the national language and lingua franca of Pakistan, with 100 million speakers worldwide. Due to resource scarcity, limited work has been done in the domain of Urdu speech recognition. In this paper, collection of Urdu speech corpus and development of Urdu speech recognition system is presented. Urdu LVCSR is developed using 300 hours of read speech data with a vocabulary size of 199K words. Microphone speech is recorded from 1671 Urdu and Punjabi speakers in both indoor and outdoor environments. Different acoustic modeling techniques such as Gaussian Mixture Models based Hidden Markov Models (GMM-HMM), Time Delay Neural Networks (TDNN), Long-Short Term Memory (LSTM) and Bidirectional Long-Short Term Memory (BLSTM) networks are investigated. Cross entropy and Lattice Free Maximum Mutual Information (LF-MMI) objective functions are employed during acoustic modeling. In addition, Recurrent Neural Network Language Model (RNNLM) is also being used for re-scoring. Developed speech recognition system has been evaluated on 9.5 hours of collected test data and a minimum Word Error Rate (%WER) of 13.50% is achieved.