Author: Ying ZhangPublisher:
ISBN:
Category :
Languages : en
Pages :
Book Description
Recurrent Neural Networks (RNNs), which has the attractive properties of modelling sequences, has been dominant in speech field in the recent decades. Convolutional Neural Networks (CNNs) has been shown as an alternative to model sequences because of its capacity of reducing spectral variations and modeling spectral correlations in acoustic features for automatic speech recognition (ASR). Recent work suggests that complex numbers could be used as a richer feature representation than spectrum which may benefit the speech related tasks. In the thesis, we first cover the basic concepts in machine learning, building blocks of deep learning and discuss the popular methods that are capable of doing sequence-to-sequence modelling, specially convolutional neural networks, which is famous as a class of feed-forward nets. We then present two research work related to sequence-to-sequence modelling on speech. We introduce a new approach to address speech recognition with convolutional neural networks which shows the comparable results with their recurrent neural networks counterpart. In addition, we present a new model taking advantage of the representation in the complex domain and define complex convolutions, complex batch-normalization, complex weight initialization strategies. The new model results in state-of-the-art of speech spectrum prediction in a convolutional recurrent setting.
