Investigating the scarce data and resources problem for speech recognition using transfer learning and data augmentation

Authors: Fahrettin Gökgöz, Mahmoud Hashem

Abstract:

We investigate the effect of the transfer learning procedure on e2e Automatic Speech Recognition systems using a limited amount of data. We use a DeepSpeech inspired base-line in our experiments and based on different transfer learning techniques. Our experimental results indicate the benefit of the augmented progressive transfer method in minimizing the over-fitting and improving the accuracy.