DNN Based Speech Enhancement for Unseen Noises Using Monte Carlo Dropout

Abstract

In this work, we propose the use of dropout as a Bayesian estimator for increasing the generalizability of a deep neural network (DNN) for speech enhancement. By using Monte Carlo (MC) dropout, we explore whether the DNN can accomplish better enhancement in unseen noisy conditions. Two DNNs are trained on speech corrupted with five different noises at three SNRs, one using conventional dropout and other with MC dropout and tested on speech with unseen noises. Speech samples are obtained from the TIMIT database and noises from NOISEX-92. In another experiment, we train five DNN models separately on speech corrupted with five different noises, at three SNRs. The model precision estimated using MC dropout is used as a proxy for squared error to dynamically select the best of the DNN models based on their performance on each frame of test data. The first set of experiments aims at improving the performance of an existing DNN with conventional dropout for unseen noises, by replacing the conventional dropout with MC dropout. The second set of experiments aims at finding an optimal way of choosing the best DNN model for de-noising when multiple noise-specific DNN models are available, for unseen noisy conditions.