Abstract:
Existing hearing aids often handle acoustic feedback, noise, and reverberation separately, thus limiting the upper performance bound of hearing aids. This paper proposes a two-stage end-to-end learning approach for joint optimization of feedback suppression, noise reduction, and dereverberation. Specially, a great number of paired data using simulated closed-loop systems in noisy and reverberant environments is generated, and these paired data were then used to train the proposed two-stage deep-learning models jointly. Finally, the pretrained models were inserted into the closed-loop system of a hearing aid to online process feedback, noise, and reverberation simultaneously in a low-latency mode. Experiments using two recorded feedback paths to evaluate the performance of the proposed approach were conducted, and experimental results showed that the proposed approach achieved the best performance in terms of PESQ, ESTOI, and frequency-weighted SNR when compared with the traditional approach, separate training models, and single-stage model.