French summary and English abstract of the present thesis, each accompanied by five keywords, as required by Doctoral School 3D.
La presente these aborde le probleme de la reconnaissance automatique de la parole pour le malgache, langue austronesienne classee a faibles ressources, parlee par plus de vingt-cinq millions de locuteurs et caracterisee par au moins dix-huit groupes dialectaux ainsi que par une diglossie marquee avec le francais. Bien que le malgache soit nominalement couvert par les principaux modeles multilingues recents (Whisper, MMS, Wav2Vec2-BERT, XLS-R), les performances en transcription en mode zero-shot restent insuffisantes pour un deploiement, et aucune technique d'adaptation prise isolement ne resout le probleme. La these propose un cadre methodologique unifie, denomme PROVA-MG (Probabilistic Routing Of Variational Alignments for Malagasy), qui aborde la reconnaissance automatique a faibles ressources comme un systeme probabiliste integrant des elements de preuve heterogenes sous calibration explicite et sous contraintes explicites. Cinq composants interconnectes structurent le cadre: la co-inference de trois modeles de fondation traites comme des posteriors distincts; l'integration d'un a priori typologique austronesien dans le decodage; un algorithme EM variationnel contraint pour le pseudo-etiquetage iteratif avec des conditions explicites de non-effondrement; une decomposition a variables latentes du dialecte, du registre et de l'alternance codique; et une tete de calibration post-hoc partagee. Le cadre s'appuie sur un substrat de donnees construit selon une logique licence-first. Une procedure de validation en huit phases avec gates mesurables soutient les pretentions de la these. Les resultats attendus, dont la confirmation est en cours, portent sur la performance globale, l'equite dialectale, la qualite de la calibration et la convergence empirique de la boucle EM contrainte.
Mots cles: reconnaissance automatique de la parole, langues a faibles ressources, malgache, modeles de fondation, EM variationnel contraint.
The present thesis addresses the problem of automatic speech recognition for Malagasy, an Austronesian language classified as low‑resource, spoken by more than twenty‑five million speakers, characterized by at least eighteen dialect groups and by a pronounced diglossia with French. Although Malagasy is nominally covered by the principal recent multilingual models (Whisper, MMS, Wav2Vec2‑BERT, XLS‑R), zero‑shot transcription quality remains below deployment thresholds, and no single adaptation technique resolves the gap. The thesis proposes a unified methodological framework, PROVA‑MG (Probabilistic Routing Of Variational Alignments for Malagasy), that approaches low‑resource ASR as a probabilistic system aggregating heterogeneous evidence under explicit calibration and explicit constraints. Five interlocking components structure the framework: co‑inference of three foundation models treated as distinct posteriors; integration of an Austronesian typological prior into decoding; a constrained variational EM algorithm for iterative pseudo‑labeling with explicit non‑collapse conditions; a latent‑variable decomposition of dialect, register, and code‑switching; and a shared post‑hoc calibration head. The framework rests on a data substrate constructed under a license‑first regime. An eight‑phase validation procedure with measurable gates supports the claims of the thesis. The expected results, whose confirmation is in progress, address headline performance, dialect equity, calibration quality, and the empirical convergence of the constrained EM loop.
Keywords: automatic speech recognition, low‑resource languages, Malagasy, foundation models, constrained variational EM.