gmovernight/COS760-ASR-Topic-Modeling-Group1

This is our project: Optimizing Topic Modeling from Speech: Evaluating ASR and Topic Model Combinations for Setswana Podcasts # COS760-ASR-Topic-Modeling-Group1 ## Optimizing Topic Modeling from Speech: Evaluating ASR and Topic Model Combinations for Setswana Podcasts **A Study on How ASR Transcription Quality Affects Topic Coherence in Low-Resource African Languages** ### Team Members (Group 1) - **Unarine Netshifhefhe** - **Tinashe Austin** - **Ruan Carlinsky** ### Github Repository Link: github.com ### Project Overview This research investigates which combination of ASR model and topic modeling technique yields the most coherent topics when processing Setswana podcast transcripts. The study systematically evaluates how different ASR systems paired with topic modeling methods affect the coherence of topics extracted from African language speech data. **Research Question:** Which combination of ASR model and topic modeling technique yields the most coherent topics for Setswana podcasts? ### Background & Motivation Most ASR models perform poorly on African languages like Setswana, significantly affecting downstream tasks like topic modeling that depend on clean text. This project addresses a critical gap in African language NLP by being the first to explore how ASR transcription quality impacts topic coherence in low-resource African languages. ### Methodology #### Three-Stage Approach: 1. **Data Preparation**: Audio standardization and preprocessing 2. **ASR Transcription**: Multiple ASR system evaluation 3. **Topic Modeling**: Comparative analysis of topic modeling techniques #### ASR Systems Evaluated: - **OpenAI Whisper**: Strong performance on low-resource languages and noisy audio - **Lelapa AI**: Optimized specifically for African languages - **wav2vec 2.0**: Self-supervised model robust to small datasets #### Topic Modeling Techniques: - **Latent Dirichlet Allocation (LDA)**: Traditional probabilistic approach - **BERTopic**: Modern transformer-based approach with better noise resilience ### Datasets **Primary Data …