Early Reflections Based Speech Enhancement
Jianfei Chen, Xihong H. Wu, Tianshu Qu · 2021
Reverberation is generally considered harmful to speech intelligibility and will cause degradation to speech related tasks. However, we propose a framework taking advantage of the early reflections (ER), which is part of reverberation, to tackle the speech enhancement problem in this work. First, a fully convolutional neural network (FCNN) is introduced to estimate the direction of arrivals (DOA) of direct sound (DS) and a few ERs. Then multiple beam signals can be attained according to the DOAs with beamforming techniques. Finally, the complex spectrum of these signals are fed into a deep neural network (DNN), which helps find an effective pattern to combine each signal to generate higher quality speech. We demonstrate the effectiveness of the proposed framework with experimental simulations, which show ERs can get higher quality speech signal, especially in severe noise conditions.