eduzhai > Applied Sciences > Engineering >

Self-supervised Learning for Speech Enhancement

  • Save

... pages left unread,continue reading

Document pages: 6 pages

Abstract: Supervised learning for single-channel speech enhancement requires carefullylabeled training examples where the noisy mixture is input into the network andthe network is trained to produce an output close to the ideal target. To relaxthe conditions on the training data, we consider the task of training speechenhancement networks in a self-supervised manner. We first use a limitedtraining set of clean speech sounds and learn a latent representation byautoencoding on their magnitude spectrograms. We then autoencode on speechmixtures recorded in noisy environments and train the resulting autoencoder toshare a latent representation with the clean examples. We show that using thistraining schema, we can now map noisy speech to its clean version using anetwork that is autonomously trainable without requiring labeled trainingexamples or human intervention.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×