Automatic Urdu Speech Recognition using Hidden Markov Model

Abstract

Pattern Recognition and Machine Learning has played a key role in the development of Automatic Speech Recognition (ASR). Most of the ASR research has been done using English Language as English is most widely spoken language in the World. However, there has been done very less research work generally in Non-English languages and particularly in Urdu language due to the interest of researchers and non-availability of resources for this language. There is no standard dataset available for Urdu language due to the lack of awareness and research and interest of the people, although Urdu and Hindi is the national language of Pakistan and India. In this research work an Urdu ASR has been developed using CMU Sphinx Toolkit. We have used a medium scale vocabulary (250) of Urdu with 250 isolated words uttered by 10 different speakers. The isolated words in the dataset are the most spoken words of Urdu language. Mel Frequency Cepstral Coefficients (MFCC) is used for feature extraction. Carnegie Mellon University (CMU) Sphinx Toolkit is used for Classification, which is based upon state of the art statistical technique Hidden Markov Model (HMM). Sphinxtrain is used for Training of the Acoustic Model. CMUCLTK is used for the development of Language Model and Pronunciation Dictionary. This work created Acoustic Model, Language Model and Dictionary for Urdu Words, which can be used in other Urdu Isolated Words Applications. This research work has achieved some promising results as compared to the previous results available in the literature on this dataset. The experiment is performed on the 100-word to compare the results with the previous work and then all the 250-words are used in another experiment.