Experimental Study on Speech Synthesis Using Advanced Neural Network Models

In this paper, the authors conduct an experimental work on neural networks Text-to-Speech (TTS), aiming to facilitate an appropriate understanding of current research and future tendencies in the field. Four cascading systems to generate an English Speech Synthesis model are used and compared. We focus on the key components in neural TTS, including text processing, acoustic-phonetics models, and speech generators - vocoders. This study is useful for industry practitioners and academic researchers interested in TTS. The experimental setup considers the following TTS models: Tacotron2 and Fastspeech2 acoustic models, both of which are used with MelGAN and Multiband MelGAN vocoders. The parameters' models are trained and tuned on the English LJSpeech dataset. Performance metrics used are WER objective and MOS subjective metrics. The mean opinion score (MOS) as subjective metrics is based on naturalness and intelligibility measures. The Results show that, in our case, the best performance is achieved using Fastspeech2 combining with Multi Band-Melgan model achieved/Comprehensibility of 4.25 and Naturalness of 4 which is close to performance of commercial systems.