Advancing Speech Enhancement with Generative Adversarial Network-Autoencoder: A Robust Adversarial Autoencoder Approach

In day-to-day life, the speech signals are often noisy and distorted by background noise. These signals are not suitable for use in different audio-operated applications directly, as they are distorted. The use of these noisy voice signals can degrade the performance of the speech communication system. There are a huge number of applications nowadays that we use for various purposes, which utilize voice as input. Our study focuses on speech enhancement, which involves a combination of Generative Adversarial Networks (GAN) and Autoencoders (AE). The required features are extracted by using the MFCC algorithm from the MUSAN dataset. The features extracted with MFCC are paired samples of clean and noisy speech. The main architecture is a combination of GAN and AE. The Generator is trained to reconstruct clean speech features from noisy speech signal inputs. On the other hand, the discriminator is trained to tell the difference between real clean samples and samples that are generated by the generator. The adversarial training approach continuously improves the performance of the generator to produce good-quality and more intelligent speech. The MUSAN dataset used for the experiment contains data of noisy speech. As a result, we note that the model performs very well and shows robustness across multiple types of noise samples. The AE is used for feature reconstruction, and the GAN for generating fake samples. This combination of GAN and AE resulted in a good solution for speech enhancement in a noisy and distorted environment.