Turkish Speech Recognition Techniques and Applications of Recurrent Units (LSTM and GRU)

Abstract

A typical solution of Automatic Speech Recognition (ASR) problems is realized by feature extraction, feature classification, acoustic modeling and language modeling steps. In classification and modeling steps, Deep Learning Methods have become popular and give more successful recognition results than conventional methods. In this study, an application for solving ASR problem in Turkish Language has been developed. The data sets and studies related to Turkish Language ASR problem are examined. Language models in the ASR problems of agglutative language groups such as Turkish, Finnish and Hungarian are examined. Subword based model is chosen in order not to decrease recognition performance and prevent large vocabulary. The recogniton performance is increased by Deep Learning Methods called Long Short Term Memory (LSTM) Neural Networks and Gated Recurrent Unit (GRU) in the classification and acoustic modeling steps. The recognition performances of systems including LSTM and GRU are compared with the the previous studies using traditional methods and Deep Neural Networks. When the results were evaluated, it is seen that LSTM and GRU based Speech Recognizers performs better than the recognizers with previous methods. Final Word Error Rate (WER) values were obtained for LSTM and GRU as 10,65% and 11,25%, respectively. GRU based systems have similar performance when compared to LSTM based systems. However, it has been observed that the training periods are short. Computation times are 73.518 and 61.020 seconds respectively. The study gave detailed information about the applicability of the latest methods to Turkish ASR research and applications.