Adversarial Text-to-Speech for low-resource languages

In this paper we propose a new method for training adversarial text-to-speech (TTS) models for low-resource languages using auxiliary data. Specifically, we modify the MelGAN (Kumar et al., 2019) architecture to achieve better performance in Arabic speech generation, exploring multiple additional datasets and architectural choices, which involved extra discriminators designed to exploit high-frequency similarities between languages. In our evaluation, we used subjective human evaluation, MOS-Mean Opinion Score, and a novel quantitative metric, the Fréchet Wav2Vec Distance, which we found to be well correlated with MOS. Both subjectively and quantitatively, our method outperformed the standard MelGAN model.