Feature Transformation for Fast and Efficient Learning in Near-Field Source Localization

This work develops an efficient and fast learning method for near-field acoustic source localization using the spherical harmonics (SH) feature transformation. The SH features are derived through the SH decomposition of the microphone array recordings. However, these SH features are often impaired by noise, interference, and reverberation, hindering localization accuracy. In this context, we proposed a feature transformation that leverages the signal subspace of the SH decomposed signals. The feature transformation reduces the irregularities in decomposed SH parameter in the noisy and reverberate condition. Therefore, the proposed subspace based feature captures significant directional and range-dependent cues for localization and enhances the training performance and accuracy with fewer epochs. The efficacy of this fast learning approach is demonstrated using convolutional neural network (CNN) training to map the input features to localization classes. The performance of the proposed approach is evaluated through exhaustive simulation and experiments and compared with the previous methods.