Deep learning meets measurement error: the hidden challenge in binaural sound source localisation

In the field of binaural sound source localisation, deep learning is often used to make predictions of a sound source’s direction. Typically, models are trained and evaluated using data synthesised from head related transfer function measurements, with the same set of measurements being used in the training and testing datasets. This work verifies whether it is valid to do so by training and testing convolutional neural networks on binaural audio made from nine different sets of measurements of the same head simulator. It is found that there is very poor cross-performance, and the models do not generalise well to unseen measurement datasets, leading to the conclusion that this approach to evaluating deep binaural sound source localisation models is not valid.