A Study on Sampling of STFT Modifications in Time and Frequency Domains for DNN-Based Speech Dereverberation

We investigate the effects of time and frequency sampling on short-time Fourier transform modifications to be used for speech dereverberation based on deep neural networks (DNNs). We first show that by adopting a linear activation function at the output layer and globally normalizing the target features into zero mean and unit variance, better performances can be obtained than existing DNN approaches. Then we show that the quality of dereverberated speech could be degraded with denser sampling in time for longer reverberation times, even at the price of increased computational complexities, requiring an adaptive time sampling strategy. On the other hand, the difference between the unwrapped phases of reverberant and anechoic speech becomes negligible with a dense sampling in frequency, implying a reduced speech distortion. Therefore, there is a great potential to enhance DNN based acoustic signal processing if the conventional sampling strategy can be carefully adjusted.