Improving Turkish Telephone Speech Recognition with Data Augmentation and Out of Domain Data

Abstract

In this paper, we investigate the effects of data augmentation and adding out of domain data on Turkish spontaneous speech recognition. We apply different acoustic model training techniques including Gaussian Mixture Models, Deep Neural Network and Time Delay Neural Network to Babel Turkish spontaneous telephone speech data. We find that Time Delay Neural Network with iVectors based acoustic model performs the best result. We demonstrate the effect of data augmentation by adding speed and volume perturbation applied data in training. We investigate the effect of increasing acoustic model training data by including two call center data. We increase training data by adding about 100 hours of modified out of domain broadcast data. We also examine the effect of neural network based language modeling techniques like Recurrent Neural Network language models.