Generative Adversarial Networks for Speech Enhancement

Speech is the fundamental form of communication among humans. Speech Enhancement is a technique that helps in improving the quality of speech. The most common method used for enhancing speech is the removal of background noise. However, speech is now being used as an input for text-based applications (such as Speech-to-Text), conferencing systems (such as Microsoft Teams), etc. A baseline GAN model is constructed to find out the inefficiencies that are present in the current GAN based approaches for speech enhancement. We propose an improved lightweight novel GAN model for speech enhancement that uses the self-attention layer adapted from non-local attention coupled with the convolutional layers in the discriminator network to improve its performance. The proposed model is evaluated on an independent and unseen test set. The audio generated by the model confirm the viability of the model.