The Application of Learnable STRF Kernels to the 2021 Fearless Steps Phase-03 SAD Challenge.

We describe a deep-learning-based system developed for the Fearless Steps Phase-03 Speech Activity Detection (SAD) challenge. The system includes both learnable spectro-temporal receptive fields (STRFs) and unconstrained 2-dimensional convolutional kernels in the first layer. Experiments show that the inclusion of learnable STRFs in the first layer increases the system's robustness to additive noise. Additionally, we found that utilizing SpecAugment during training improves generalization on unseen data. By incorporating these enhancements and others our system achieved the best score in the official SAD challenge.