A Review on Low Complexity Deep Learning Based Techniques for Acoustic Scene Analysis

For the past few years, research in audio related technology has changed the focus to Acoustic Scene analysis using Deep Learning Techniques. The major task in Acoustic Scene analysis is to classify the recorded sound based on the labels used for training the network. Some examples of acoustic scenes include the market, beach, traffic, office, classroom etc. Acoustic Scene Analysis (ASA) or Acoustic Scene Classification (ASC) is the term used for the classification of sound based on the aforementioned situations. Detection and Classification of Acoustic Scenes and Events (DCASE) focus on this ASC by conducting the competition, which was started in the year 2013. So far, every competition has included one or more ASC challenges. In this work, historical overview of earlier studies was provided in the field of ASC before defining a generic framework for ASC. With an emphasis on low complexity deep learning, this article discusses and classifies popular techniques for data preparation, such as feature representations, feature pre-processing, data augmentation, and network architectures for ASC.