An Auditory-Based Monaural Feature for Noisy and Reverberant Speech Enhancement

The deep neural networks (DNN) based speech enhancements is a hot topic in machine learning and speech enhancement application. Even with deep neural network, it is still hard to improve the speech quality on noisy and reverberant conditions. For machine learning based system, auditory feature extraction becomes the key point in speech enhancement and recognition. In this paper, we proposed a speech enhancement framework based on an auditory-based monaural feature, which model the function of human hearing auditory system. The auditory based feature is extracted from the data passing the gammatone filter banks, which has more detail on low frequency than normal filters. Systemic tests show the better performance of the proposed auditory based monaural feature than the mel-frequency cepstral coefficients (MFCC) in noise and reverberant environment.