Customizable End-To-End Optimization Of Online Neural Network-Supported Dereverberation For Hearing Devices

Abstract

This work focuses on online dereverberation for hearing devices using the weighted prediction error (WPE) algorithm. WPE filtering requires an estimate of the target speech power spectral density (PSD). Recently deep neural networks (DNNs) have been used for this task. However, these approaches optimize the PSD estimate which only indirectly affects the WPE output, thus potentially resulting in limited dereverberation. In this paper, we propose an endto-end approach specialized for online processing, that directly optimizes the dereverberated output signal. In addition, we propose to adapt it to the needs of different types of hearing-device users by modifying the optimization target as well as the WPE algorithm characteristics used in training. We show that the proposed endto-end approach outperforms the traditional and conventional DNNsupported WPEs on a noise-free version of the WHAMR! dataset.

End-To-End Optimization of Online Neural Network-supported Two-Stage Dereverberation for Hearing Devices

“…We noticed in our previous study [25] that although the energy residing in the moderate reverberation range corresponding to the filter length was particularly suppressed when training the approach in an end-to-end fashion, residual late reverberation could still be heard at the output. A further processing stage could be dedicated to removing this residual reverberation, as increasing the length of the linear filters results in rapidly increasing computational complexity and training difficulty.…”

“…In [12], [20], the DNN is optimized with a mean-squared error (MSE) criterion on the masked output. In contrast, in our previous work [25] we proposed to use the Kullback-Leibler (KL) divergence [31]:…”

“…1) End-to-end criterion and objectives: In [25], we showed that the mismatch between the DNN-optimization criterion (7) and the dereverberation task limited the overall performance. On the other hand, we argued that using ASR as an end-toend training criterion, as is done in [24], is not necessarily the best choice in order to optimize a dereverberation algorithm for hearing-aid users.…”

“…An end-to-end procedure using ASR was also introduced in [24] to optimize a DNN used for online dereverberation. In contrast, we proposed to use a criterion directly on the output signal rather than using ASR in our previous work [25]. We argued that it was more likely to improve instrumentally predicted speech intelligibility and quality.…”

“…We show with the newly introduced metrics that this latter stage particularly benefits from strong dereverberation within the linear filter range obtained with the previous end-to-end WPE approach. Finally, we customize the presented two-stage algorithm to hearing listener classes by adapting the training target and algorithm parameters as in previous work [25].…”