Multi-MelGAN voice conversion for the creation of under-resourced child speech synthesis

Abstract

Voice conversion (VC) is an important technique for the development of text-to-speech voices in the use case of lacking speech resources. VC can convert an audio signal from a source speaker to a specific target speaker whilst maintaining the linguistic information. The benefit of VC is that you only require a small amount of target data which therefore makes it possible to build high quality text-to-speech voices using only a limited amount of speech data. In this work, we implement VC using a Melspectrogram Generatative Adversarial Network called MelGAN-VC. This technique does not require parallel data and has been proven successful on as little as 1 hour of target speech data. The aim of this work was to build child voices by modifying the original one-to-one MelGAN-VC model to a many-to-many model and determine if there is any gain in using such a model. We found that using a many-to-many model performs better than the baseline one-to-one model in terms of speaker similarity and the naturalness of the output speech when using only 24 minutes of speech data.