Prova de Doutoramento da aluna Mariana Dimas Julião Simões Coelho
Área: Engenharia Informática e de Computadores
Título da Tese: Prosody Assessment for Computer-Assisted Language Learning
Local da Prova: Anfiteatro PA-3 (Piso -1 do Pavilhão de Matemática) do IST
Data: 16/01/2026
Hora: 09h30
Abstract: Prosody has gained increasing attention in Computer-Assisted Language Learning (CALL); however, it remains underemphasized in broader language education. A major limitation of many CALL systems is their reliance on imitation-based assessment, which primarily evaluates how closely learners mimic native speakers rather than assessing their actual prosodic competence. Furthermore, these systems often fail to provide actionable feedback, leaving learners aware of their performance level but without clear guidance on how to improve. A more effective approach would assess prosody based on its functional role in communication rather than its proximity to native patterns, fostering a learner-centered methodology that better supports language acquisition. This thesis contributes to the field of L2 prosody assessment by exploring diverse methodologies while adhering to these learner-centered principles. The first part of this work investigates different methods for representing prosody, evaluating a range of acoustic-prosodic features, neural embeddings, and other fixed-length representations across various prosody-related tasks. The second part explores the feasibility of prosody assessment through comparisons with different reference sources: first, native speakers, and later, synthetic speech. By comparing TTS-generated prosody to both native and non-native productions across proficiency levels, we find that while TTS outputs often appear plausible, they lack the variability necessary to fully capture the range of natural speech. In the final part, we introduce Goodness of Prosody, an event-based system that generates symbolic prosody references from text and evaluates L2 productions using an acoustic classifier. The system's results correlate with speaker proficiency levels, providing insights into the challenges and limitations of automatic prosody assessment. These findings are thoroughly analyzed, highlighting both the potential and obstacles of this approach.