eduzhai > Applied Sciences > Engineering >

Learning to Denoise Historical Music

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: We propose an audio-to-audio neural network model that learns to denoise oldmusic recordings. Our model internally converts its input into a time-frequencyrepresentation by means of a short-time Fourier transform (STFT), and processesthe resulting complex spectrogram using a convolutional neural network. Thenetwork is trained with both reconstruction and adversarial objectives on asynthetic noisy music dataset, which is created by mixing clean music with realnoise samples extracted from quiet segments of old recordings. We evaluate ourmethod quantitatively on held-out test examples of the synthetic dataset, andqualitatively by human rating on samples of actual historical recordings. Ourresults show that the proposed method is effective in removing noise, whilepreserving the quality and details of the original music.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×