eduzhai > Applied Sciences > Engineering >

Dense CNN with Self-Attention for Time-Domain Speech Enhancement

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Speech enhancement in the time domain is becoming increasingly popular inrecent years, due to its capability to jointly enhance both the magnitude andthe phase of speech. In this work, we propose a dense convolutional network(DCN) with self-attention for speech enhancement in the time domain. DCN is anencoder and decoder based architecture with skip connections. Each layer in theencoder and the decoder comprises a dense block and an attention module. Denseblocks and attention modules help in feature extraction using a combination offeature reuse, increased network depth, and maximum context aggregation.Furthermore, we reveal previously unknown problems with a loss based on thespectral magnitude of enhanced speech. To alleviate these problems, we proposea novel loss based on magnitudes of enhanced speech and a predicted noise. Eventhough the proposed loss is based on magnitudes only, a constraint imposed bynoise prediction ensures that the loss enhances both magnitude and phase.Experimental results demonstrate that DCN trained with the proposed losssubstantially outperforms other state-of-the-art approaches to causal andnon-causal speech enhancement.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...