SAMPA: Whisper-based prosodic boundary segmentation in Brazilian Portuguese

This paper introduces SAMPA, a segmenter based on the Whisper large-v3 model designed to identify terminal prosodic boundaries in Brazilian Portuguese speech.

• Fine-tuned on manually segmented recordings from the NURC-SP dataset.
• Achieved an F1 score of 0.731 on the held-out test split and 0.796 on the MuPe-Diversidades dataset.
• Demonstrates reliance on morphosyntactic, semantic, and prosodic cues through n-gram and acoustic-visual analyses.

The model provides a competitive deep learning alternative to traditional rule-based methods for automatic prosodic segmentation in Brazilian Portuguese.