Glottalization in the light of Machine Learning – Glot-TAL

The Glot-TAL project addresses a crucial issue in phonetics and phonology: glottalization in the world's tonal languages. By combining phonetic expertise with the power of deep learning and large-scale data analysis, it will allow for significant progress in fundamental phonetic-phonological theory and linguistic typology, and will moreover open up new perspectives for Natural Language Processing.

The project pursues two major scientific objectives.

From a linguistic perspective, the project aims to explore laryngeal activity in human communication by examining the links between pitch control, glottalization and larynx position in tone production. The focus is on tonal languages such as Vietnamese, where glottalization holds phonological importance.

From a machine learning perspective, the goal is to investigate how deep learning can effectively integrate signals other than audio – specifically: electroglottography (EGG) – into statistical tools for speech analysis. Experiments will be carried out using existing corpora in the Pangloss Collection, then extended to a sample of widely different languages (reaching a new level in terms of typological coverage and amounts of data).

To achieve these ambitious challenges, the project relies on close collaboration between linguists and experts in automatic natural language processing, and is thus fully in line with trends in digital humanities. It goes beyond the development of tools for linguists in analyzing less-documented tonal languages worldwide. By reaching an unprecedented scope in terms of linguistic diversity, it will bring new perspectives to fundamental questions in linguistic research on the use of glottalization in human communication.