Ensemble learning based head-related transfer function personalization using anthropometric features

An ensemble learning based model to predict the logarithmic magnitude response of the head-related transfer function (HRTF) using anthropometric features is proposed. Subjects are first clustered based on relevant anthropometric features, then the ensemble learning algorithm is used on clustered results for predicting the log-magnitude response of HRTF. The model contains two levels of deep neural networks (DNNs). The first-level DNNs predict log-magnitude HRTFs in clustered groups, and the second-level DNN integrates estimates produced by the first-level DNNs. Experimental results show the proposed model outperforms compared DNN models in terms of the log-spectral distortion (LSD) measure with greater stability.