Importance of switch optimization criterion in Switching WPE dereverberation

Weighted prediction error (WPE) is a fundamental dereverberationmethod to predict the late reverberation component of an observedsignal based on linear prediction (LP). Recently, WPE was extendedto Switching WPE (SwWPE), which optimizes (i) multiple LP filtersand (ii) switching parameters to determine the best LP filter used foreach time-frequency bin. Conventionally, these parameters are op-timized based on the maximum likelihood (ML) criterion, but thisis not optimal in terms of signal quality, such as signal-to-distortionratio (SDR) and word error rate (WER) of automatic speech recogni-tion. We thus propose a new SwWPE processing flow that enables usto optimize switching parameters based on an arbitrary optimizationcriterion. Using oracle clean signals, we demonstrate the potentialperformance of our new approach with an SDR maximization crite-rion, revealing that it can significantly improve the SDR and WERobtained by the conventional ML-based SwWPE. This motivates usto propose new SwWPE processing in which the switching param-eters are externally estimated using a deep neural network (DNN)that is trained with an end-to-end SDR maximization criterion. Theexperimental result clearly demonstrates the improved SDR perfor-mance of the new approach compared to the conventional WPE andSwWPE.