Speech enhancement through improvised conditional generative adversarial networks

Speech enhancement works towards improvising the quality of speech through various post processing algorithms. Intelligibility enhancement along with overall perceptual quality score improvement is the main objective of many speech signal processing techniques. Generative Adversarial Networks (GAN) aims to generate a new set of data with the help of training set statistics and is seen to be impressive for enhancing the speech signals in the recent years. Though GAN's does not involve prior and posterior probability calculations, they are hard to train in general. The problem aggravates with low-data regime and hence there is a need for effective GAN mechanism. In this research work, we propose to use an improvised conditional generative adversarial network where the generator will enhance the input data that is noisy while the discriminator on the other hand embedded with improvised techniques will try to differentiate between the generator output and the database clean content with the help of GAN conditions discussed. The results of the proposed method are assessed in terms of PEAQ score and equal error rate. Experimental results from the Aurora-2 signal set proves us that the improved cGAN is very effective as compared to traditional GAN networks. We have also tested the algorithm with the subjective preference and 82.14% of the subjects were found to prefer the proposed cGAN to that obtained with other conventional methods.