http://hdl.handle.net/10386/66 Tue, 14 Apr 2026 11:12:31 GMT 2026-04-14T11:12:31Z http://hdl.handle.net/10386/5380 Development of a Sepedi-English code-switching automatic speech recognition system using connectionist temporal classification Phaladi, Amanda Speech technology includes several approaches and technologies that allow ma- chines to engage with spoken language, which include spoken dialog systems and automatic speech recognition. The end-to-end (E2E) techniques, such as Connec- tionist Temporal Classification (CTC) and attention-based methods, dominate Auto- matic Spdeech Recognition (ASR) system development. However, these methodolo- gies have primarily advanced in research for high-resourced languages with exten- sive speech datasets, leaving low-resource languages relatively underserved. The efficacy of the CTC method specifically for Sepedi, a low-resource language, remains uncertain. This study addresses this gap by developing and evaluating an automatic speech recognition (ASR) system for Sepedi-English code-switched speech. Utilizing the Se- pedi Prompted Code Switching (SPCS) corpus and applying the CTC approach, we implemented an E2E ASR system. We rigorously evaluated the system’s performance across various parameters using both the National Centre for Human Language Tech- nology (NCHLT) Sepedi test corpus and the Sepedi Prompted Code Switching corpus. Our findings demonstrate promising results overall. However, the system faced challenges in accurately recognizing speech from the Sepedi NCHLT test corpus. This study shows the importance of adapting advanced ASR techniques to suit the linguistic characteristics and data limitations of low-resource languages. Addressing these challenges is crucial for expanding the applicability of speech technology to diverse linguistic contexts, ultimately facilitating broader accessibility and usability of ASR systems worldwide. Thesis (M.Sc. (Computer Science)) -- University of Limpopo, 2025 Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10386/5380 2025-01-01T00:00:00Z http://hdl.handle.net/10386/5379 Analysing the explainability of credit scoring machine learning models using Shapley Additive Explanations approach Thoka, Merriam Ramakgahlele In recent years, machine learning models have gained popularity in credit scoring applications due to their ability to handle large volumes of data and capture complex patterns. However, the lack of transparency and interpretability in these models raises concerns regarding their trustworthiness and fairness. This study aims to address this matter by employing the Shapley Additive Explanations (SHAP) approach to analyse the explainability of credit scoring machine learning models. The lending club dataset, a comprehensive collection of loan applications and associated attributes, is utilized for this analysis. The methodology involves training and evaluating various credit scoring models, including Random Forest, XGBoost, and CatBoost, and generating SHAP values to quantify the importance of input features in the prediction process. The results reveal valuable insights into the factors influencing credit scoring decisions and provide a holistic understanding of the models’ behaviour. By utilizing SHAP explanations, we gain interpretability and can identify features that significantly impact the credit scoring outcomes. This knowledge can help stakeholders, including lenders and regulators, make informed decisions and improve the transparency and accountability of credit scoring systems. The discoveries of this study advance the expanding field of explainable artificial intelligence(AI) and its application in the domain of credit risk management. By enhancing the explainability of credit scoring models, we aim to increase trust, fairness, and accountability in the lending process, ultimately shaping a more inclusive and responsible financial ecosystem. Thesis (M. Sc. (eScience)) -- University of Limpopo, 2025 Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10386/5379 2025-01-01T00:00:00Z http://hdl.handle.net/10386/5376 Development of an end-to-end automatic speech recognition system using connectionist temporal classification for the Tshivenda language Mehlape, Jonas Mosweu This study centers on creating an automatic speech recognition (ASR) system for Tshivenda, one of South Africa's under-resourced languages. Utilizing the Connectionist Temporal Classification (CTC) framework and the NCHLT speech corpus. This study focuses on developing an E2E ASR system leveraging CTC techniques for the Tshivenda language. The primary objective is to develop and evaluate an ASR system for the Tshivenda language using the CTC approach. This involves designing and training an ASR model using the NCHLT speech corpus, optimizing model performance through hyperparameter tuning (e.g., learning rate, dropout rate), and evaluating the system’s accuracy through essential metrics such as WER and training loss. The research also focuses on identifying key challenges in recognizing Tshivenda speech and proposes improvements for future work in this area. However, there are several delimitations to the scope of the study that should be considered. First, the research relies on the NCHLT speech corpus, which, although valuable, has limited dialectal diversity and does not fully represent all regional variations of Tshivenda. Additionally, the model was primarily trained on clean speech data, and as such, it does not extensively address the challenges of handling noisy environments or spontaneous speech. Furthermore, while the study focuses on a CTC-based deep learning model, it does not explore the integration of external language models, such as transformer-based models, which could further enhance performance. Finally, due to hardware limitations, the model was trained for 30 epochs, which may have constrained the model's ability to reach its optimal performance, potentially impacting the accuracy of the final system. The model's performance was assessed over 30 epochs using essential metrics, including Word Error Rate (WER), training loss, and validation loss. The top-performing model achieved a final WER of 0.3934, highlighting notable advancements in Tshivenda speech recognition. This research highlights the promise of deep learning models in creating ASR systems for under-resourced languages, while also pointing out critical directions for future exploration. Key advancements include expanding the dataset, integrating language models, and improving the model’s resilience to noisy conditions and spontaneous speech. These steps are essential for enhancing accuracy and practical usability. The study contributes to the broader mission of promoting language preservation and accessibility through technological innovation Thesis (M. Sc. (Computer Science)) -- University of Limpopo, 2025 Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10386/5376 2025-01-01T00:00:00Z http://hdl.handle.net/10386/5363 Developing a code-mixed sentiment analysis model for Xitsonga-English music review Nkuna, Blessing Sentiment analysis is an essential natural language processing technique for monitoring online discussions about brands, products, and services. Traditionally focused on monolingual data, sentiment analysis has now expanded to include code-mixed texts, reflecting the growing use of multiple languages within single sentences on social media. This dissertation addresses the gap in sentiment analysis for code-mixed data by developing a Long Short-Term Memory (LSTM) classifier for Xitsonga-English comments extracted from YouTube music reviews. This research aims to design and implement a sentiment analysis model tailored for Xitsonga-English code-mixed texts, evaluating its performance against traditional monolingual sentiment analysis methods. This includes collecting a substantial dataset of Xitsonga-English comments, determining their polarity, developing an LSTM classifier, and assessing its accuracy, precision, recall, and F1-score. Data collection involved scraping 1 998 Xitsonga-English comments from a Xitsonga YouTube channel, cleaning and tokenizing the comments for analysis. Sentiments were defined and categorized into positive, negative, and neutral classes based on specific criteria, with dictionaries developed for both Xitsonga and English lexicons. These lexicons were used to label the comments, facilitating the creation of training data for the LSTM model. Additionally, a word embedding matrix was developed using Word2Vec, capturing semantic similarities between words. The LSTM classifier's architecture included embedding layers initialized with pre-trained word embeddings, two LSTM layers for sequence processing, and a dense output layer for sentiment classification. Despite efforts to address overfitting through regularization and model adjustments, the final LSTM model did not perform as expected on the validation and test datasets, highlighting challenges in generalizing sentiment classification for the collected dataset. To address this, a stacking classifier combining Random Forest, Support Vector Machine, Gradient Boosting, and Logistic Regression was developed and compared with the LSTM model. The stacking classifier showed better generalization on unseen data, indicating its robustness for sentiment analysis tasks in code-mixed contexts. The results highlight the challenges and potential solutions in developing robust sentiment analysis models for code-mixed languages, contributing valuable insights to the domain of natural language processing. Thesis (M. Sc. (Computer Science)) -- University of Limpopo, 2025 Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10386/5363 2025-01-01T00:00:00Z