ridwanbello/african-clinical-asr-proposal

# African Clinical ASR Proposal Experiments This repository contains code, configurations, result tables, and documentation for my PhD proposal on efficient domain adaptation of pretrained ASR models for African-accented clinical speech recognition. ## Dissertation Goal The goal of this dissertation is to analyze and improve efficient adaptation strategies for pretrained ASR models on African-accented clinical speech, with emphasis on recognition performance, computational efficiency, and general-domain preservation. ## Research Questions ### RQ1: Error Analysis What error patterns do state-of-the-art ASR models exhibit on African-accented clinical speech, and how do these patterns vary across clinical and general-domain speech? ### RQ2: PEFT vs Full Fine-Tuning How do parameter-efficient fine-tuning methods compare with full fine-tuning for adapting Whisper-based ASR models to African-accented clinical speech in terms of recognition performance, computational efficiency, and trainable parameter cost? ### RQ3: Domain-Aware Adaptation What domain-aware adaptation strategy can best balance specialization to African-accented clinical speech with preservation of general-domain ASR capability? ## Repository Structure - `shared/`: shared utilities for data loading, normalization, metrics, and plotting - `rq1_error_analysis/`: scripts, configs, and results for RQ1 - `rq2_peft_comparison/`: scripts, configs, and results for RQ2 - `rq3_domain_aware_peft/`: scripts, configs, and results for RQ3 - `tables/`: proposal-ready result tables - `docs/`: experiment logs, data paths, weekly notes, and proposal mapping ## Current Deadlines - Proposal submission: August 19, 2026 - Proposal defense: October 1, 2026