Exploring the Impact of Data Augmentation Techniques on Emotional Speech Recognition

Automatic Speech Recognition (ASR) systems are known to perform poorly in recognition of atypical speech, which certainly includes emotional speech. The labor-intensive nature of developing comprehensive speech databases for training purposes has led to the widespread adoption of synthetic speech data generation, utilizing existing natural speech, in numerous research studies. This paper examines the impact of standard data augmentation techniques, including pitch shifting, time stretching and volume adjustment on the accuracy of isolated-word recognizer of emotional speech. The experiments carried out indicate a significant increase in accuracy achieved through a single augmentation based on pitch-shifting for all emotional states.