کنفرانس مهندسی برق ایران

صفحه اصلی / سی و یکمین کنفرانس بین المللی مهندسی برق

Improving CycleGAN-VC2 Voice Conversion by Learning MCD-Based Evaluation and Optimization

نویسندگان :

Majid Behdad¹ Davood Gharavian²

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی

کلمات کلیدی :

CycleGAN-VC،perceptual evaluation،perceptual optimization،MetricGAN،Mel-Cepstral distance،speech quality Assessment،Nisqa tool

چکیده :

Abstract—Nowadays’ voice conversion systems that convert source speakers to target speakers in a speech utterance, have various applications, and improving their quality is very important. One method that still has not attracted enough attention for the VC quality improvements is to concentrate on the optimization of the discriminators of a GAN-based VC System. In this paper, we chose Cycle-GAN-VC2 as the baseline to implement a modified version of Mel-scale human hearing-related objective evaluation metric, Modified Mel-Cepstral Distance (MMCD) to help the discriminators to better learn to judge between real and fake data. We developed and implemented our new metric MMCD that is between 0 and 1 to use it in discriminators’ loss functions. The main goal is to force the discriminators to learn the MMCD metric behavior in its judgements; while in conventional CycleGAN-VC2, discriminators work as the classifiers that decide which data is real and which one is fake without any attention to perceptual references and measures like MCD score that can be varied continuously from zero to one. Experimental results show improvements in the quality of output speech versus MCD measure despite that the training of our baseline VC system is based on a set of non-parallel data, and don’t use any time-alignment in training process. So, in parallel VC systems more improvements could be anticipated.