Unsupervised feature learning on monaural DOA estimation using convolutional deep belief networks

In recent years, deep learning approaches have gained significant interest as a way of building hierarchical representations from unlabeled data. Additionally, in the field of sound direction-of-arrival (DOA) estimation, the binaural features like interaural time or phase difference and interaural level difference, or monaural cues like spectral peaks and notches are often used to estimate sound DOA. Although these binaural or monaural cues successfully explained human sound DOA ability, its accuracy and extent are all limited to the human knowledge on feature extracting methods. In this paper, we are interested in applying a deep learning approach to monaural sound localization based on monaural spectral cues. A convolutional deep belief network is applied to monaural auditory spectrograms processed by a computational auditory model to learning monaural features automatically. The learned features are then regressed using a support vector regression (SVR) model for sound DOA estimation tasks. Moreover, our feature representations learned from unlabeled monaural auditory spectrograms are then compared to the well-known binaural features. The results indicate that our monaural model shows reasonable performance at the task of 3D DOA estimation.