Ensemble Learning Approach With Class Rotation for Three-Dimensional Classification on Direction-of-Arrival Estimation

Over the past decade, neural networks have been widely used for direction-of-arrival (DoA) estimation owing to their high accuracy in noisy and reverberant environments. Classes of singular-model classifiers generally correspond to discretized DoA candidate angles, which in the case of a three-dimensional estimation, are bounded by the grid derived from uniform sampling over the unit sphere. Motivated by this, we propose an ensemble learning approach for classification tasks to improve estimation accuracy, as the ensemble criterion outperforms the individual criterion. First, individual networks that make up the ensemble differ slightly by grid rotation according to the Euler rotation theorem to complement discrete directional information. Score fusion was performed in the spherical harmonic domain for a more stable ensemble classification because the grid rotation also involves an angular mismatch. Moreover, to achieve a more accurate DoA estimation, interpolation over the fused scores was performed. Performance analysis and a comparison of state-of-the-art parametric and deep learning-based methods in several acoustic situations were conducted to determine the accuracy of the ensemble and to analyze the gradual angular error reduction as a function of noise and reverberation levels as more networks were added.