Time Delay Neural Network for Myanmar Automatic Speech Recognition

Abstract

Time Delay Neural Network (TDNN) contains in neural network architectures. In Automatic Speech Recognition, TDNN is strong possibility in context modeling and recognizes phonemes and acoustic features, independent of position in time. There are many techniques have been applied for improving Myanmar speech processing. TDNN based acoustic model for Myanmar ASR in this paper. Myanmar language is a low resource language and no precollected data is available. A larger dataset and lexicon than our previous work are applied in this experiment. The speech corpus contains three domains: Names, Web News data and Daily conversational data. The size of the corpus is 77 Hrs and 2 Mins and 11 Secs and include 233 female speakers and 97 male speakers. The performance of TDNN for Myanmar ASR is shown by comparing with Gaussian Mixture Model (GMM) as a baseline system, Deep Neural Network (DNN) and Convolutional Neural Network (CNN). Experiments evaluation is used 2 test data: TestSet1, web news and TestSet2, recorded conversational data. The experimental results show that TDNN outperforms GMM-HMM, DNN and CNN.