Improving Low Resource Turkish Speech Recognition with Data Augmentation and TTS

Abstract

One of the major problems faced by speech recognition researchers is the lack of data. In this paper, our objective is to compare alternative solutions to lack of data. Some experiments are conducted with very limited training data to see the effects of data augmentation and speech synthesis on speech recognition. Speed and volume perturbations are applied in this study. Besides data augmentation, synthetic speech is generated by using two different speech synthesis methods. In first speech synthesis approach, Google Translate Text to Speech (gTTS) is used as speech synthesizer. In second speech synthesis approach, an end-to-end Turkish TTS system is trained by us. Finally, we examined the effects of all these alternative methods on speech recognition for low resource languages. Our results demonstrate that some data augmentation or speech synthesis techniques work well to improve speech recognition for low resource languages. In this study, 14.8% relative Word Error Ratio (WER) improvement is obtained by using combination of augmented and synthetic data.