A Novel End-to-End Turkish Text-to-Speech (TTS) System via Deep Learning

Abstract

Text-to-Speech (TTS) systems have made strides but creating natural-sounding human voices remains challenging. Existing methods rely on noncomprehensive models with only one-layer nonlinear transformations, which are less effective for processing complex data such as speech, images, and video. To overcome this, deep learning (DL)-based solutions have been proposed for TTS but require a large amount of training data. Unfortunately, there is no available corpus for Turkish TTS, unlike English, which has ample resources. To address this, our study focused on developing a Turkish speech synthesis system using a DL approach. We obtained a large corpus from a male speaker and proposed a Tacotron 2 + HiFi-GAN structure for the TTS system. Real users rated the quality of synthesized speech as 4.49 using Mean Opinion Score (MOS). Additionally, MOS-Listening Quality Objective evaluated the speech quality objectively, obtaining a score of 4.32. The speech waveform inference time was determined by a real-time factor, with 1 s of speech data synthesized in 0.92 s. To the best of our knowledge, these findings represent the first documented deep learning and HiFi-GAN-based TTS system for Turkish TTS.