In this thesis, we develop an open-source Slovenian text-to-speech
(TTS) system capable of incorporating emotions and paralinguistic vocal
events, such as laughter, interjections, and breathing, into synthesized speech.
Since appropriately annotated data for the Slovenian language is scarce, we
first develop a pipeline for automatically annotating speech recordings with
vocal events and emotions. We compare conventional approaches to audio
event and emotion detection with modern approaches based on large language
models (LLMs). Using the resulting annotations, we adapt the open-source
Slovenian TTS model CosyVoice 3 and evaluate its ability to generate more
expressive and natural-sounding speech.
|
