Wavelet Scattering Based Deep Gated Recurrent Units for Binaural Acoustic Scenes Classification

A Gated Recurrent Units (GRUs) for Acoustic Scenes Classification (ASC) with wavelet time scattering is proposed. The neural networks using a combination of multiple GRU layers in the Neural Network enhance ASC’s accuracy. Due to the advantage of time sequence modeling, the GRUs can perform better class-based discrimination and mapping features. The results on the TUT Acoustic Scenes evaluation dataset demonstrates that the proposed model performs better in ASC than the bidirectional Long short-term memory (LSTM) layer and its counterpart LSTM. The average class-wise accuracy has shown around 23.7% improvement with the proposed GRUs network. Compared to other highest performance LSTMs based method, GRUs offer 3.6% and 4.8% improvement in the classification accuracy and F1-Score, respectively.