Representation Learning for Automatic Speech Recognition: A Review of Speech-to-Text Methods

Abstract

Representation learning has emerged as a promising approach to overcoming the limitations of discriminative repre sentations from the raw speech signal. In this review, we cover a range of speech-to-text methods that employ representation learning, including deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer-based models. The advantages and limitations of each approach are described, as well as recent advances in pretraining techniques such as contrastive predictive coding (CPC) and masked language modelling (MLM). The reviewed papers are divided according to their novelty, their approaches and their type of representation learning models.