The increasing importance of spatial audio technologies has demonstrated the need and importance of correctly adapting to the individual characteristics of the human auditory system, and illustrates the crucial need for humanoid localization systems for testing these technologies. To this end, this paper introduces a novel feature analysis and selection approach for binaural localization and builds a probabilistic localization mapping model, especially useful for the vertical dimension localization. The approach uses the mutual information as a metric to evaluate the most significant frequencies of the interaural phase difference and interaural level difference. Then, by using the random forest algorithm and embedding the mutual information as a feature selection criteria, the feature selection procedures are encoded with the training of the localization mapping. The trained mapping model is capable of using interaural features more efficiently, and, because of the multiple-tree-based model structure, the localization model shows robust performance to noise and interference. By integrating the direct path relative transfer function estimation, we propose to devise a novel localization approach that has improved performance in the presence of noise and reverberation. The proposed mapping model is compared with the state-of-the-art manifold learning procedure in different acoustical configurations, and a more accurate and robust output can be observed.
Robust Binaural Localization of a Target Sound Source by Combining Spectral Source Models and Deep Neural Networks
