Exploring mini-batch sample selection strategies for deep learning based speech recognition

This thesis aims to propose mini-batch sample selection strategies for deep learning based speech recognition systems. Deep learning based speech recognition systems became more prevalent and state-of-the-art system for speech recognition domain with the popularity and success of deep learning architectures. Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) RNN are widely and successfully utilized for the applications of speech recognition. Mini-batch gradient descent algorithm is generally accepted algorithm for training deep learning based speech recognition systems. Mini-batch gradient descent algorithm is a successful algorithm, but one of the main problems in mini-batch gradient descent is that the training samples are selected randomly for each mini-batch. In this thesis, mini-batch sample selection strategies are proposed to improve speech recognition accuracy of deep learning based speech recognition systems. Proposed strategies use meta features of speech corpuses, i.e. gender and accent features. Three types of sample selection strategies are proposed, i.e. gender adjusted strategies, accent adjusted strategies, and hybrid strategies that combine gender and accent adjusted strategies. The experimental results show that proposed strategies are beneficial for improving performance of deep learning based speech recognition systems.