Abstract:

Being able to isolate a target speech signal from background signals is of direct importance for telephony, hands-free communication and audio surveillance, and it is also critical as a pre-processing step in applications such as voice activity detection, automatic speaker identification, and most importantly automatic speech recognition (ASR) in challenging environments. While speech enhancement and separation methods originally did not rely on training, there has recently been an explosion in the use of machine learning based methods that exploit large amounts of training data. This tutorial will present a broad overview of these methods, analyzing the insights that can be gained from the pre-deep-learning era of graphical modeling and NMF approaches, then diving into an in-depth presentation of recent deep learning approaches encompassing single-channel methods, multi-channel methods, and new directions.