gms | German Medical Science

Reverberant environments pose challenges for people with hearing impairments, especially for users of cochlear implants. Thus algorithms to enhance speech in reverberant conditions have been developed, and one such algorithm is Weighted Prediction Error (WPE) [1], which requires an estimate of the anechoic speech power spectral density (PSD). In recent work [2], we developed an online capable version of WPE where the PSD estimate is computed by a deep neural network (DNN). Rather than giving a straight anechoic speech PSD, the DNN is trained to optimize the WPE output with respect to the desired criterion. We term this approach E2Ep-WPE.The WPE algorithm proved very efficient at suppressing early reflections and moderate reverberation, but is not able to remove the late reverberant tail of the room impulse response. Thus, we further modified the algorithm to include a post-processing stage using a second DNN. We label this enhancement algorithm E2Ep-WPE+DNN-PF.This contribution reports an initial subjective assessment of the proposed E2Ep-WPE and E2Ep-WPE+DNN-PF, using a MUSHRA-like presentation to normal-hearing subjects. This comparison allows us to evaluate the possible benefit of the post-filter, which incurs additional computational complexity but no additional delay. We also include in our benchmark GaGNet [3], a DNN-based state-of-the-art enhancement algorithm and successor of the 2021 DNS challenge winning approach.