reachmaheshkandagatla/Afrikaans_ASR

# Afrikaans ASR Research Framework This repository provides a reproducible framework for training and evaluating an Afrikaans automatic speech recognition (ASR) model using OpenAI Whisper. It was developed from an initial Colab workflow and reorganized into a structured Python project for research, experimentation, and future model iteration. The framework focuses on adapting `openai/whisper-large-v3` to Afrikaans speech through parameter-efficient fine-tuning with LoRA. The pipeline separates data preparation, dataset validation, model training, evaluation, and Hugging Face Hub publication so that each research phase can be inspected and repeated independently. ## Research Motivation Afrikaans remains underrepresented in many modern ASR benchmarks and production-ready speech systems. While multilingual foundation models such as Whisper provide a strong starting point, domain-specific adaptation is often required to improve transcription quality for local accents, recording conditions, vocabulary, and speech patterns. This project supports research into: - low-resource and underrepresented-language ASR adaptation, - Whisper fine-tuning for Afrikaans transcription, - LoRA-based training under limited compute budgets, - reproducible dataset preprocessing for speech models, - word error rate (WER) evaluation of Afrikaans ASR outputs. ## Objectives The main objectives of this framework are: 1. Convert raw Afrikaans audio/transcript datasets into Whisper-compatible features. 2. Validate mapped data before GPU training to avoid expensive training failures. 3. Fine-tune Whisper Large v3 with LoRA for Afrikaans transcription. 4. Track model quality using WER during evaluation. 5. Save and publish trained adapters and processors for reuse. 6. Provide a clean foundation for future experiments, ablations, and dataset expansion. ## Methodology The workflow is divided into three research phases. ### Phase 1: Dataset Mapping Raw Hugging Face `datasets` splits are con …