Prosodic Phrasing and Interrogative Melodic Contour in Czech Text-to-Speech Syntheses

Recent advances in synthetic speech have substantially improved the naturalness of TTS (text-to-speech) systems, however, challenges regarding prosody modelling persist, potentially impacting cognitive speech processing. This thesis investigates prosodic tendencies between neural text-to-speech systems and human read speech, focusing on prosodic phrasing and intonation contour in questions. The recordings were annotated in Praat and acoustic parameters including f0 contour, number of syllables, and prosodic boundary structure were extracted. The results revealed stronger dependence of neural TTS speech on textual punctuation for prosodic phrasing, as well as differences in intonation patterns across question types. A perceptual evaluation conducted with human listeners showed that synthetic speech was reliably recognized and rated as less natural, particularly when its prosodic patterns diverged from those observed in human speech. These findings highlight limitations in current prosody modelling for Czech language and underscore the importance of context evaluation and greater variability in training data for neural speech synthesis systems. Keywords: text-to-speech, prosody, prosodic phrase, intonation, neural voices in Czech