Speech Dereverberation Based on Integrated Deep and Ensemble Learning

摘要

Reverberation, which is generally caused by sound reflections from walls,ceilings, and floors, can result in severe performance degradations of acousticapplications. Due to a complicated combination of attenuation and time-delayeffects, the reverberation property is difficult to characterize, and itremains a challenging task to effectively retrieve the anechoic speech signalsfrom reverberation ones. In the present study, we proposed a novel integrateddeep and ensemble learning (IDEL) algorithm for speech dereverberation. TheIDEL algorithm consists of offline and online phases. In the offline phase, wetrain multiple dereverberation models, each aiming to precisely dereverb speechsignals in a particular acoustic environment; then a unified fusion function isestimated that aims to integrate the information of multiple dereverberationmodels. In the online phase, an input utterance is first processed by each ofthe dereverberation models. The outputs of all models are integratedaccordingly to generate the final anechoic signal. We evaluated IDEL ondesigned acoustic environments, including both matched and mismatchedconditions of the training and testing data. Experimental results confirm thatthe proposed IDEL algorithm outperforms single deep-neural-network-baseddereverberation model with the same model architecture and training data.