A neural network-supported two-stage algorithm for lightweight dereverberation on hearing devices

A two-stage lightweight online dereverberation algorithm for hearing devices is presented in this paper. The approach combines a multi-channel multi-frame linear filter with single-channel single-frame post-filter. Both components rely on power spectral density (PSD) estimates provided by deep neural networks (DNNs). By deriving new metrics analyzing the performance various time ranges, we confirm that directly optimizing criterion at output of filtering stage results more efficient as compared to placing DNN optimize PSD estimation. More concretely, show training end-to-end helps further remove reverberation range accessible filter, thus increasing \textit{early-to-moderate} ratio. We argue and demonstrate it can then be well combined post-filtering efficiently suppress residual late reverberation, thereby \textit{early-to-final} This proposed two procedure shown both very effective terms computational demands, e.g. recent state-of-the-art approaches. Furthermore, system adapted needs different types hearing-device users controlling amount reduction early reflections.