Beyond sentence prosody: paragraph intonation in Dutch

Area of Science:

• Computational Linguistics
• Speech Science
• Phonetics

Background:

• Synthesized text intonation models are crucial for natural-sounding speech.
• Understanding prosodic variations in text is essential for improving speech synthesis.
• Previous models often lack detailed analysis of positional effects within discourse.

Purpose of the Study:

• To develop a partial model for synthesized text intonation in single-phrase utterances.
• To investigate the influence of utterance position within a paragraph on prosody.
• To enhance the naturalness of rule-based intonation in automatic text-to-speech (TTS) conversion.

Main Methods:

• Formulating a model based on intonational characteristics of short utterances.
• Analyzing prosodic variations across different positions within a paragraph.
• Perceptual evaluation of the formulated intonation model.

Main Results:

• Positional effects on sentence prosody are primarily observed at the beginning of utterances.
• The developed model effectively captures key findings regarding positional intonation.
• Perceptual evaluation confirmed the model's ability to improve naturalness.

Conclusions:

• Paragraph position significantly influences synthesized intonation, particularly at utterance beginnings.
• The proposed partial model offers a viable approach to improving rule-based TTS intonation.
• Further research can extend this model to more complex utterances and discourse structures.