SAMPA: Transformer-based segmentation of prosodic boundaries in Brazilian Portuguese

Poster presented at the 16th Lisbon Machine Learning School (LxMLS 2026), Lisbon, Portugal. URL https://lxmls.it.pt/2026 This poster presents SAMPA, a Whisper-based approach for automatic prosodic boundary segmentation in Brazilian Portuguese (BP). Fine-tuned on the NURC-Sampa corpus and evaluated across both in-distribution and out-of-distribution (MuPe-Diversidades) datasets, SAMPA demonstrates that fine-tuning a speech recognition model to jointly transcribe and mark terminal prosodic boundaries outperforms standard baselines and traditional classifiers, advancing deep learning-based prosody modeling for BP.