Subspace Gaussian Mixture Model for Continuous Urdu Speech Recognition using Kaldi

Automatic Speech Recognition Systems (ASR) have significantly improved in recent years, where deep learning is playing an important role in the development of end to end ASR's. ASR is the task of converting spoken language into computer readable text. ASRs are becoming ever more prevalent way to interact with technology, thereby significantly closing the gap in terms of how humans interact with computers, making it more natural. Urdu is an under resourced language, for which training such a system requires a huge amount of data that is not readily available. In this paper we present improvements to the architecture of a statistical automatic speech recognition system for which the components involved in a statistical ASR have been explored in great detail. We also present the results on various statistical models that are trained for Urdu language. We choose the Kaldi toolkit for training the Urdu ASR using approximately 100 hours of transcribed data. The refined Subspace Gaussian Model gives a word error rate of 9% on the test set.