Quantization-Error-Free Soft Label for 2D Sound Source Localization

Linfeng Feng, Xiao-Lei Zhang, Xuelong Li

Abstract: One of the state-of-the-art direction of arrival (DOA) estimation techniques is formulated as a classification problem using deep learning. However, it inherently suffers from quantization errors during the classification formulation. This weakness is further amplified in two-dimensional (2D) sound source localization (SSL). To address this limitation in 2D SSL, this paper aims to develop a quantization-error-free training objective, named Unbiased Label Distribution (ULD), along with a corresponding decoding scheme for the predicted distribution. The key idea is to use multiple adjacent classes jointly to eliminate quantization error. Experimental results show that the proposed algorithm significantly breaks the quantization error limit when the classification model achieves high accuracy. It also demonstrates strong robustness in low signal-to-noise ratio, high reverberation, and far-field environments.