Deep representation learning for speech recognition
Równicka, Joanna Małgorzata
Representation learning is a fundamental ingredient of deep learning. However, learning a good representation is a challenging task. For speech recognition, such a representation should contain the information needed to perform well in this task. A robust representation should also be reusable, hence it should capture the structure of the data. Interpretability is another desired characteristic. In this thesis we strive to learn an optimal deep representation for speech recognition using feed-forward Neural Networks (NNs) with different connectivity patterns. First and foremost, we aim to improve the robustness of the acoustic models. We use attribute-aware and adaptive training strategies to model the underlying factors of variation related to the speakers and the acoustic conditions. We focus on low-latency and real-time decoding scenarios. We explore different utterance summaries (referred to as utterance embeddings), capturing various sources of speech variability, and we seek to optimise speaker adaptive training (SAT) with control networks acting on the embeddings. We also propose a multi-scale CNN layer, to learn factorised representations. The proposed multi-scale approach also tackles the computational and memory efficiency. We also present a number of different approaches as an attempt to better understand learned representations. First, with a controlled design, we aim to assess the role of individual components of deep CNN acoustic models. Next, with saliency maps, we evaluate the importance of each input feature with respect to the classification criterion. Then, we propose to evaluate layer-wise and model-wise learned representations in different diagnostic verification tasks (speaker and acoustic condition verification). We propose a deep CNN model as the embedding extractor, merging the information learned at different layers in the network. Similarly, we perform the analyses for the embeddings used in SAT-DNNs to gain more insight. For the multi-scale models, we also show how to compare learned representations (and assess their robustness) with a metric invariant to affine transformations.