eduzhai > Applied Sciences > Engineering >

Deep Noise Suppression With Non-Intrusive PESQNet Supervision Enabling the Use of Real Training Data

  • KanKan
  • (4) Download
  • 20210428
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Data-driven speech enhancement employing deep neural networks (DNNs) canprovide state-of-the-art performance even in the presence of non-stationarynoise. During the training process, most of the speech enhancement neuralnetworks are trained in a fully supervised way with losses requiring noisyspeech to be synthesized by clean speech and additive noise. However, in a realimplementation, only the noisy speech mixture is available, which leads to thequestion, how such data could be advantageously employed in training. In thiswork, we propose an end-to-end non-intrusive PESQNet DNN which estimatesperceptual evaluation of speech quality (PESQ) scores, allowing areference-free loss for real data. As a further novelty, we combine the PESQNetloss with denoising and dereverberation loss terms, and train a complexmask-based fully convolutional recurrent neural network (FCRN) in a "weakly "supervised way, each training cycle employing some synthetic data, some realdata, and again synthetic data to keep the PESQNet up-to-date. In a subjectivelistening test, our proposed framework outperforms the Interspeech 2021 DeepNoise Suppression (DNS) Challenge baseline overall by 0.09 MOS points and inparticular by 0.45 background noise MOS points.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×