Towards Reliable Prosody Detection in Teaching Practical Phonetics: A Sustainable Digital Approach.
- 1The Department of Modern Languages and Communication, National University of Science and Technology MISIS, Moscow, Russian Federation.
- 2Liberal Arts Department, American University of the Middle East, Egaila, Kuwait, 54200, Kuwait.
Abstract
This research focuses on providing quality education to specialists in the field of linguistic studies, phonology, theoretical and practical phonetics by creating an innovative automated service which can ease the linguists endeavour to transcribe and examine suprasegmental features of human speech prosody. The authors attempted to design a docker-based architecture for prosody labelling. The pre-processing stage compiled 47 manually marked recordings, ten of which were annotated at clip level. The data-processing stage produced a corpus of 882 denoised, labelled clips via librosa, noisereduce, and Label Studio. The model selection stage compared 4909 scikit-learn configurations and 120 CNN configurations. The strongest classical approach (Random Forest) achieved an accuracy of 0.373, whereas the best CNN scored 0.455 against a majority-class baseline of 0.316. Per-class analysis showed that the classical models performed poorly across most intonation categories, assigning 94% of predictions to the two most frequent classes. These results indicate early-stage prototype performance rather than a validated evaluation and are not yet sufficient to support reliable automatic intonation marking. Having recognised the prototype limitations, the researchers have scheduled a roadmap for improving the tool. The suggested model applies the principles of machine learning to automatically generate prosodic analysis that extends beyond individual sound segments and reflects suprasegmental aspects of prosody including rhythm, intonation, and phrasal stress.
Related
- Evaluating the introduction of a practical thyroid ultrasound teaching session for medical students
- Ten simple rules for teaching practical AI to health data teams
- Perceptions of residents and graduates regarding the theoretical-practical teaching model in orthopedics.
- Transforming Practical Veterinary Anatomy Teaching (2013-2026): From Demonstration to Team-Based Learning.
- Practical versus Theoretical Teaching of Abnormal Physical Examination Findings to Medical Students: A Randomized Controlled Trial
BibTeX
@article{tolstykh2026towards,
title = {{Towards Reliable Prosody Detection in Teaching Practical Phonetics: A Sustainable Digital Approach.}},
author = {Olesya Tolstykh and Alexey Svatov and Tamara Oshchepkova},
journal = {F1000Research},
year = {2026},
month = oct,
doi = {10.12688/f1000research.178913.3},
url = {https://europepmc.org/article/MED/42892806},
note = {CC BY},
}
RIS
TY - JOUR
TI - Towards Reliable Prosody Detection in Teaching Practical Phonetics: A Sustainable Digital Approach.
AU - Olesya Tolstykh
AU - Alexey Svatov
AU - Tamara Oshchepkova
PY - 2026
DA - 2026/10/10
JO - F1000Research
DO - 10.12688/f1000research.178913.3
UR - https://europepmc.org/article/MED/42892806
AB - This research focuses on providing quality education to specialists in the field of linguistic studies, phonology, theoretical and practical phonetics by creating an innovative automated service which can ease the linguists endeavour to transcribe and examine suprasegmental features of human speech prosody. The authors attempted to design a docker-based architecture for prosody labelling. The pre-processing stage compiled 47 manually marked recordings, ten of which were annotated at clip level. The data-processing stage produced a corpus of 882 denoised, labelled clips via librosa, noisereduce, and Label Studio. The model selection stage compared 4909 scikit-learn configurations and 120 CNN configurations. The strongest classical approach (Random Forest) achieved an accuracy of 0.373, whereas the best CNN scored 0.455 against a majority-class baseline of 0.316. Per-class analysis showed that the classical models performed poorly across most intonation categories, assigning 94% of predictions to the two most frequent classes. These results indicate early-stage prototype performance rather than a validated evaluation and are not yet sufficient to support reliable automatic intonation marking. Having recognised the prototype limitations, the researchers have scheduled a roadmap for improving the tool. The suggested model applies the principles of machine learning to automatically generate prosodic analysis that extends beyond individual sound segments and reflects suprasegmental aspects of prosody including rhythm, intonation, and phrasal stress.
SN - 2046-1402
ER -