Trainable windows for SincNet architecture

Abstract SincNet architecture has shown significant benefits over traditional Convolutional Neural Networks (CNN), especially for speaker recognition applications. comprises parameterized Sinc functions as filters in the first layer followed by convolutional layers. Although is compact nature and offers top-level understanding of features extracted, effect window function used not thoroughly addressed yet. Hamming Hann are popularly default time-localized windows to reduce spectral leakage. Hence, a comprehensive investigation 28 different windowing on towards task using TIMIT dataset was performed this work. Additionally, “trainable” were configured with tunable parameters characterize performance. The paper benchmarks terms bandwidth, side-lobe suppression, leakage filter banks employed architecture. Trainable Gaussian Cosine-Sum exhibited relative improvement 41.46% 82.11% sentence level classification error rate when