Robust Interaural Time Difference Estimation Based on Convolutional Neural Network

Hong Liu, Peipei Yuan, Bing Yang, Lulu Wu

Abstract: This paper proposes a novel cross correlation function (CCF) extraction method based on convolutional neural network for time difference of arrival (TDOA) estimation or further direction of arrival (DOA) estimation. CNN is utilized to learn the relationship between the cross correlation localization features and the pre-processed waveform signal which may include not only the source signal but also the background noise and reverberation. In contrast to many previous sound source localization approaches, the proposed method focuses on the spatial feature extraction. Two kind of outputs, grouped or encoded CCF, are designed to capture the implicit tendency of location information. The experimental results demonstrate that the proposed method outperforms the conventional TDOA estimation methods under environments with different levels of noise and reverberation.