Voice Conversion and Spoofed Voice Detection from Parallel English and Urdu Corpus using Cyclic GANs

With the advent of Generative Adversarial Networks (GANs), the fake news epidemic is booming; which not only encompasses pictures and videos but also audio. This is a big issue in an automatic speech verification (ASV) devices allowing anyone to steal an identity from a database of users. We aim to address this issue for a database of speaker utterances in the Urdu language by a two-fold solution. First, we will describe a Cyclic GAN based one-to-one conversion method that can generate speech from given speaker to a target voice bi-directionally. Cyclic GANs have much more strong mapping capabilities than ordinary GANs due to the property of Cyclic consistency loss. This framework ensures that given sufficient training data, generated output is very similar to the input. Furthermore, adversarial examples generated by the model are used for spoofed voice detection. We will use a Gradient Boosting method to learn to distinguish the voice utterances of various speakers that are stored in a database from the adversarial examples. For the testing of English language, we used the VCTK dataset and for the Urdu language, we used Urdu speech recordings containing a single word utterance from each speaker. This is tested for male → male, male → female, female → male and female → female voice conversions. The results obtained from learning from the adversarial examples are optimistic but more data and efforts are needed to make it usable into practical systems that can support speech verification at large scale.