Exploring Machine Learning perspectives for electroglottographic signals - Archive ouverte HAL Accéder directement au contenu
Rapport Année : 2023

Exploring Machine Learning perspectives for electroglottographic signals

Exploration des perspectives d'apprentissage automatique pour les signaux électroglottographiques

Résumé

Studying and preserving under-resourced and endangered languages, for which few resources are available, is an arduous endeavour, which involves tremendously time-consuming tasks. In addition, for best results down the line (when the datasets are exploited for a variety of purposes), consistency in annotation work is highly desirable, but not easy to ensure in a manual workflow. High hopes are placed in innovative uses of speech recognition tools to facilitate and accelerate linguistic tasks. Automatic Speech Recognition (ASR) is making remarkable progress thanks to the advent of deep neural networks (DNNs). A breakthrough is now possible: machine learning tools have improved to a point where they can effectively help to perform linguistic annotation tasks such as automatic transcription of audio recordings, automatic glossing of texts, and automatic word discovery. However, beyond that, there exists a considerable space for computer-assisted exploration and analysis of phonetic and phonological properties of languages. In this context, we attempt to explore the applicability of a neural network for phonetic/phonological analysis of an audio and electroglottographic corpus that has been previously processed manually. The objective of this work is to test the capabilities of neural networks to learn and reproduce specific strategies and principles for the estimation and manual verification of two phonetic parameters, namely fundamental frequency and glottal open quotient, which are acoustic correlates of pitch and phonation type (both of which constitute relevant dimensions of linguistic tone in the target language). In order to evaluate this learning process, a comparison of automatic and manual results is carried out. This pilot study uses a recently collected and manually analyzed corpus1 of the Kim Thuong dialect of Muong, a Vietic language that has a phonetically complex tonal in which one tone involves a lapse into creaky voice (M.-C. Nguyen, 2021). The acoustic and electroglottographic signals were recorded simultaneously. The manual annotation was based on the acoustic signal. The measurement of fundamental frequency and glottal open quotient was based on peak detection on the derivative of electroglottographic signal (also known as the DEGG signal) using the semi-automatic Peakdet script running on Matlab. Data from twenty speakers with twelve minimal sets plus three minimal pairs recorded within a carrier sentence amount to a total of five hours of recordings (an average of 18 minutes per speaker). The results point to the presence of confounders, which (paradoxically) include the use of a carrier sentence: placing the target items in the same phonetic context stabilizes their tonal realizations, but the artificial neural network is biased towards easier predictions. But analysis of the results has benefits for being more explicit on the goals of analysis of electroglottographic signals, offering pointers for further work.
Fichier principal
Vignette du fichier
Exploring_Machine_Learning_perspectives_for_electroglottographic_signals___Minh_Châu_NGUYÊN___livrable_CLD2025.pdf (1.38 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04081199 , version 1 (25-04-2023)
hal-04081199 , version 2 (07-06-2023)

Licence

Paternité - Pas d'utilisation commerciale - Partage selon les Conditions Initiales

Identifiants

  • HAL Id : hal-04081199 , version 2

Citer

Minh-Châu Nguyễn. Exploring Machine Learning perspectives for electroglottographic signals. LIG (Laboratoire informatique de Grenoble). 2023. ⟨hal-04081199v2⟩
83 Consultations
66 Téléchargements

Partager

Gmail Facebook X LinkedIn More