Experiments of speech enhancement based on deep neural networks in far field scenarios.

Abstract

Efficient speech enhancement techniques are essential to improve quality and intelligibility of speech signals and reliability in voice recognition applications. One way to reduce word error rate is through powerful noise reduction algorithms. This paper is intended to provide speech enhancement experiments based on deep neural networks (DNN) in far field scenarios under noisy environments. In the present work, it is investigated the use of DNNs as a front-end for automatic speech recognition (ASR). Objective metrics are used to investigate its effectiveness and results are compared against several popular speech enhancement algorithms. It is shown that DNN provides better results under all noisy and SNR conditions.