MuphulusiDzivhani/Automatic-Speech-Recognition-ASR-and-Topic-Modeling-for-African-Languages

COS802 Project – Automatic Speech Recognition (ASR) and Topic Modeling for African Languages # 📊 COS 802 – Automatic Speech Recognition (ASR) and Topic Modeling for African Languages This repository includes a research project on Automatic Speech Recognition (ASR) for Xitsonga podcast data, achieving **85.3% accuracy** with comprehensive Word Error Rate assessment and linguistic analysis. ## 👨‍🎓 Author Credentials **Muphulusi Dzivhani (u18069682)** University of Pretoria COS 802 Research Project Module ## 📖 About This Project The objective of this project is to bridge the digital gap for African Languages through an ASR pipeline for Xitsonga, a low-resource language in South Africa with 4.5 million native speakers. This research fine-tunes OpenAI's Whisper model using 117 Nalibali Xitsonga podcast episodes and achieves production-grade accuracy. ## 🎯 Research Questions - To what accuracy can Whisper accommodate natural podcast speech in Xitsonga? - What is the Word Error Rate for African language data using fine-tuned models? - How accurate is transcription for language-specific features like click consonants? - What is the most effective preprocessing pipeline for Xitsonga audio? - What are the key ASR implementation challenges for African languages? ## 🛠️ Implementation of the Technical Component ### ASR Solution - **Model**: Fine-tuned OpenAI Whisper-small - **Platform**: Hugging Face Transformers + Google Colab - **Audio Preprocessing**: Librosa - **Quantitative Metrics**: Custom WER calculation + visualization ### Advantages - **117 files** processed with 100% success rate - **85.3% accuracy** achieved on Xitsonga speech - Comprehensive Word Error Rate assessment - Linguistic analysis of Xitsonga features - Fully replicable automated pipeline ## 📊 Results ### Performance Metrics - **Word Error Rate**: 14.7% across 117 Xitsonga podcast files - **Accuracy**: 85.3% (commercial-grade performance) - **Processing Effective Rate**: 100% (117/117 files processed) - **Click Consonants Recognition**: ~85% - **Processing Speed**: ~60 seconds per f …