Audio post-production for film involves the manipulation of large amounts of audio data. There is a need for the automation of many organization tasks currently performed manually by sound engineers, such as grouping and renaming multiple audio recordings. Here, we present a method to classify such sound files in two categories, ambient recordings and single-source sounds. Automating these classification tasks requires a deep learning model capable of answering questions about the nature of each sound recording based on specific features. This study focuses on the relevant features for this type of audio classification and the design of one possible model. In addition, an evaluation of the model is presented, resulting in high accuracy, precision and recall values for audio classification.
Authors: Guillermo G. Peeters (Queen Mary University of London) and Joshua D. Reiss (Queen Mary University of London)