Neural networks for text-to-speech in lithuanian

This paper describes the process of training a neural network for text-to-speech synthesis in Lithuanian language. It also includes the steps for creating custom datasets used in TTS systems. For training purposes, we are using Rayhane Mamah implementation of Google's Tacotron 2 system. Three single-speaker Lithuanian language datasets were created and used in TTS model training. All datasets differ in size and quality. The first one includes 102 hours of speech, but is poorely processed. The second one has 23.5 hours of speech and is carefully revised for quality. The third one has both properties: 92 hours of speech data and it is revised. Finally, we compare the results. Text-to-speech systems are widely used in fields such as entertainment, assistance tool development for people with disabilities, communicational aid system development and many more. Only concatenative TTS systems are known to be developed for Lithuanian language. TTS systems that use neural networks demonstrate better quality speech synthesis, so it is important to research the possibilities of implementing such systems for Lithuanian language