← PROVA‑MG overview
Front matter  /  Figures and tables

List of Figures and Tables

Indicative inventory of the figures and tables that accompany the present thesis. Numbering follows the chapter-figure convention used throughout the manuscript.

Part One

List of figures

NumberCaption
Figure 1.1Noisy-channel decomposition of speech recognition: acoustic likelihood and language-model prior.
Figure 1.2Schematic of an HMM-GMM acoustic model with three-state phone topology.
Figure 1.3WFST decoding graph: composition of H, C, L and G transducers.
Figure 1.4Hybrid DNN-HMM training pipeline using forced alignment.
Figure 1.5Comparison of CTC, attention-based encoder-decoder, and RNN-Transducer architectures.
Figure 1.6Conformer block: macaron feed-forward, multi-head self-attention, and convolution module.
Figure 1.7Self-supervised pre-training pipeline of Wav2Vec 2.0 with contrastive loss.
Figure 1.8Masked-prediction objective in HuBERT and its iterative target-refinement loop.
Figure 2.1Cross-lingual speech representations: language coverage of XLS-R, MMS, and Wav2Vec2-BERT.
Figure 2.2Whisper architecture overview and weak-supervision corpus composition.
Figure 2.3Adapter and LoRA parameter-efficient adaptation schematic.
Figure 2.4Pseudo-labeling loop with confidence filtering.
Figure 2.5License-first data pipeline: nine stages from collection to release.
Figure 3.1Block diagram of the PROVA-MG routing function.
Figure 3.2Graphical model of latent dialect, register, and code-switching with observed acoustic and transcript variables.
Figure 3.3Constrained variational EM loop and the four constraint families.
Figure 3.4Calibration head: features and post-hoc temperature / isotonic mapping.
Figure 3.5Integration of the five components with the existing license-first pipeline.
Figure 4.1Reliability diagram before and after calibration on the FLEURS-MG development set.
Figure 4.2Stratified WER across four Malagasy dialect regions.
Figure 4.3Convergence curves of constrained vs unconstrained iterative pseudo-labeling.
Figure 4.4Compute cost vs WER for teacher ensemble and distilled student.
Figure 4.5Routing decision distribution by agreement bucket.
Part Two

List of tables

NumberCaption
Table 1.1Comparison of four ASR paradigms: acoustic model, language-model integration, typical data, key reference.
Table 1.2Self-supervised speech models: pre-training objective, target, languages.
Table 2.1Cross-lingual foundation models considered: hours, languages, license, role in PROVA-MG.
Table 2.2Evaluation datasets for Malagasy ASR: hours, speakers, dialects, license.
Table 2.3Open methodological problems and their formal status.
Table 3.1Routing function: feature catalogue and decision outcomes.
Table 3.2Constraint set C of the variational EM with admissibility conditions.
Table 3.3Calibration features: source, semantics, and expected sign of effect.
Table 4.1Headline WER on FLEURS-MG, CV-MG, and in-domain test sets.
Table 4.2Component-leave-one-out ablation.
Table 4.3Routing vs averaging vs voting comparison.
Table 4.4Typological prior leave-one-language-out study.
Table 4.5Constrained-EM leave-one-constraint-out study.
Table 4.6Expected Calibration Error before and after fitting.
Table 4.7Dialect-stratified WER (Merina, Betsimisaraka, Sakalava, Antandroy).
Table 4.8Code-switching density buckets and WER per bucket.
Table 4.9Compute cost comparison: per-utterance and per-hour figures.
Table 4.10Publication plan: contributions, target venues, status.