A unified methodology for Automatic Speech Recognition that aggregates heterogeneous foundation models, typological priors, and a large licensed unlabeled corpus under explicit calibration and explicit constraints, together with the validation procedure that supports its claims.
Malagasy is nominally covered by every modern multilingual ASR release, yet zero‑shot quality is poor and no single foundation model, supervision regime, or adaptation technique improves it on its own.
PROVA‑MG — Probabilistic Routing Of Variational Alignments for Malagasy — treats low‑resource ASR as a probabilistic system that aggregates heterogeneous evidence. Three foundation models (Whisper‑large‑v3, MMS, Wav2Vec2‑BERT) are retained as distinct posteriors and their disagreement is used as a routing signal; an Austronesian typological prior constrains decoding where a Malagasy‑only language model cannot; iterative pseudo‑labeling is formalized as a constrained variational EM procedure with explicit non‑collapse conditions; dialect, register, and code‑switching are unified as latent variables; and a calibration head, trained on the small supervised set, ensures that every confidence feeding into the system is trustworthy.
The four chapters below carry the argument: the foundational state of the art, the multilingual and low‑resource state of the art together with the open methodological gaps, the PROVA‑MG methodology and its integration with the existing data pipeline, and the validation procedure, expected applications, and limitations.
A guided climb for readers who know only school algebra: one new symbol at a time, from variables and probability up to the novel constrained variational‑EM formula, each step read aloud in plain English.
Open →Hook, terms, context, problem statement, hypotheses, objective, methodology, expected results, plan of the thesis.
Open →The four paradigms on which modern ASR is built: HMM‑GMM, hybrid DNN‑HMM, end‑to‑end objectives, and self‑supervised representations.
Open →Cross‑lingual self‑supervision, weak supervision, parameter‑efficient adaptation, pseudo‑labeling, speech‑LLMs, and license‑first data, with the open methodological problems identified.
Open →Three teachers as distinct posteriors, an Austronesian prior, constrained variational EM, latent dialect, register and code‑switching, and a calibration head; integration with the existing pipeline.
Open →Prototype implementation, experimental matrix (placeholder), discussion, perspectives, positioning with respect to prior art, publication plan, eight‑phase validation.
Open →Recapitulation of the argument, main findings as expected, reminder of novelty, limits, and open perspectives for further work.
Open →