Deep Neural Network-Based Noise Estimation for Robust ASR in Dual-Microphone Smartphones

Abstract

The performance of many noise-robust automatic speech recognition (ASR) methods, such as vector Taylor series (VTS) feature compensation, heavily depends on an estimation of the noise that contaminates speech. Therefore, providing accurate noise estimates for this kind of methods is crucial as well as a challenge. In this paper we investigate the use of deep neural networks (DNNs) to perform noise estimation in dual-microphone smartphones. Thanks to the powerful regression capabilities of DNNs, accurate noise estimates can be obtained by just using simple features as well as exploiting the power level difference (PLD) between the two microphones of the smartphone when employed in close-talk conditions. This is confirmed by our word recognition results on the AURORA2-2C (AURORA2 - 2 Channels - Conversational Position) database by largely outperforming single- and dual-channel noise estimation algorithms from the state-of-the-art when used together with a VTS feature compensation method.