Deeply supervised curriculum learning for deep neural network-based sound source localization

Abstract

Deep neural network (DNN) has made impressive progress in sound source localization (SSL) tasks with the hard n-hot labels that represent specific directions-of-arrivals (DOAs). However, recent study suggested soft DOA labels, considering the correlations between targets and nearby DOAs. In this study, to effectively train a DNN using soft labels, we propose deeply supervised curriculum learning (DSCL) by adopting the two techniques for the DNN, deep supervision (DS) and curriculum learning (CL). We train a DNN to solve SSL problems progressing from easier to harder, expecting the DNN would gradually reduce the angular region of the target DOAs. It is gained by various resolution soft targets for the different DNN layers to deeply supervise the DNN, while increasing the angular selectivity of the targets from the early to late stages of training by CL. Proposed method was verified on datasets with multi-speakers, and exceeded the hard-label methods with great improvements.