Dereverberation of Relative Harmonic Coefficients via CNNs for Acoustic Source DOA estimation

Relative Harmonic Coefficients (RHCs) are a promising audio descriptor for Direction of Arrival (DOA) estimation but are vulnerable to noise and reverberation. We introduce RHC-ED, a convolutional encoder-decoder architecture that processes noisy and reverberant RHCs, restoring their ideal properties by suppressing unwanted artifacts. Using stacked CNNs, RHC-ED compresses and reconstructs RHCs for improved DOA estimation. Experiments across diverse acoustic conditions confirm RHC-ED’s effectiveness in reducing estimation errors and outperforming recent state-of-the-art methods for source localization, especially using first-order spherical harmonics.