Classification of Spatial Audio Location and Content Using Convolutional Neural Networks

This paper investigates the use of Convolutional Neural Networks for spatial audio classification. In contrast to traditional methods that use hand-engineered features and algorithms, we show that a Convolutional Network in combination with generic preprocessing can give good results and allows for specialization to challenging conditions. The method can adapt to e.g. different source distances and microphone arrays, as well as estimate both spatial location and audio content type jointly. For example, with typical single-source material in a simulated reverberant room, we can achieve cross-validation accuracy of 94.3% for 40-ms frames across 16 classes (eight spatial directions, content type speech vs. music).