A Perceptually Inspired Data Augmentation Method for Noise Robust CNN Acoustic Models

摘要

Here, we present a data augmentation method that improves the robustness of convolutional neural network-based speech recognizers to additive noise. The proposed technique has its roots in the input dropout method because it discards a subset of the input features. However, instead of doing this in a completely random fashion, we introduce two simple heuristics that select the less reliable components of the spectrum of the speech signal as candidates for dropout. The first heuristic retains spectro-temporal maxima, while the second is based on a crude estimation of spectral dominance. The selected components are always retained, while the dropout step discards or retains the unselected ones in a probabilistic manner. Due to the randomness involved in dropout, the whole process may be interpreted as a data augmentation method that perturbs the data by creating new data instances from the existing ones on the fly. We evaluated the method on the Aurora-4 corpus just using the clean training data set, and we got relative word error rate reductions between 22% and 25%.