Building HMM-SGMM Continuous Automatic Speech Recognition on Myanmar Web News

Abstract

Myanmar language is a tonal and analytic language. It can be considered as an under-resourced language because of its linguistic resource availability. Therefore, speech data collection is a very challenging task in building Myanmar automatic speech recognition. Today a lot of speech data are freely available on the Internet and we can collect it easily. Therefore, in this system, we take the advantages of Internet and we use daily news from the Web in building our speech corpus. In this paper, we will present about the task of data collection, the effect of Automatic Speech Recognition (ASR) performance according to amount of training data, language model size and error analysis of the experimental result. The experiments will be developed using Hidden Markov Model (HMM) with Gaussian Mixture Model (GMM) and Subspace Gaussian Mixture Model (SGMM). As a result, using our developed 5 hours training data, this system achieves word error rate (WER) of 7.6% on close test data and 31.9% on open test data with HMM-SGMM.