Replicating Human Sound Localization with a Multi-Layer Perceptron

One of the key capabilities of the human sense of hearing is to determine the direction from which a sound is emanating, a task known as localization. This paper describes the derivation of a machine learning model which performs the same localization task: Given an audio waveform which arrives at the listener’s eardrum, determine the direction of the audio source. Head-related transfer functions (HRTFs) from the ITA-HRTF database of 48 individuals are used to train and validate this model. A series of waveforms is generated from each HRTF, representing the sound pressure level at the listener’s eardrums for various source directions. A feature vector is calculated for each waveform from acoustical properties motivated by prior literature on sound localization; these feature vectors are used to train multi-layer perceptrons (MLPs), a form of artificial neural network, to replicate the behavior of single individuals. Data from three individuals are used to optimize hyperparameters of both the feature extraction and MLP stages for model accuracy. These hyperparameters are then validated by training and analyzing models for all 48 individuals in the database. The errors produced by each model fall in a log-normal distribution. The median model is capable of identifying, with 95% confidence, the sound source direction to within 20 degrees. This result is comparable to previously-reported human capabilities and thus shows that an MLP can successfully replicate the human sense of sound localization