Abstract
This paper analyzes the generalization of speech enhancement algorithms based on deep neural networks (DNNs) with respect to (1) the chosen features, (2) the size and diversity of the training data, and (3) different network architectures. To address (1), we compare three input features, namely logarithmized noisy periodograms, noise aware training (NAT) and signal-to-noise ratio (SNR) based noise aware training (SNR-NAT). For improving the generalization over noisy periodograms, NAT appends an estimate of the noise power spectral density (PSD) to these features, whereas the proposed SNR-NAT uses the noise PSD estimate for normalization. To address (2), we train networks on the Hu noise corpus (limited size), the CHiME 3 noise corpus (limited diversity) and also propose a large and diverse dataset collected based on freely available sounds. On the one hand, we show that increasing the amount and the diversity of training data helps DNNs to generalize. On the other hand, via experimental results and an analysis using t-distributed stochastic neighbor embedding (t-SNE) we show that SNR-NAT features are robust even if the size and the diversity of the input data are limited. To address (3), we compare a fully-connected feed-forward DNN and an long short-term memory (LSTM) and show that the LSTM generalizes better for limited training data and simplistic features.
Abstract (translated)
URL
https://arxiv.org/abs/2004.03512
