Efficient Deep Learning-Base Speech Synthesis Model Study for TTS

Abstract

TTS technology using deep learning is very similar to human speech. In order to make such TTS technology, it receives text, creates a spectrogram is Text-to-Mel, and turns the spectrogram is Text-to-Mel, and turns the spectrogram into speech in a vocoder. However, there are various Text-to-Mel and Vocoder models. We need to devise a way to analyze  and combine the various Text-to-Mel and Vocoders produce good sound quality. Therefore, in this paper, in order to convent English text to speech, a 24_hour LJ Speech Dataset that one person reads 7 books was used. And this dataset was used to train Text-to-Mel Models and vocoder models. Each learned Text-to-Mel and vocoder models were combined with each other, and the combination performance was compared for which one was the best combination and which one was the worst.