Analyzing the Regularization Technique in Automatic Speech Recognition System Using Sepedi-English Code-Switched Data

Automatic Speech Recognition (ASR) systems playa crucial role in facilitating communication across diverse linguistic contexts. However, when dealing with code-switched data, such as Sepedi- English, these systems face significant challenges due to the inherent complexity and variability of such language mixing. Dropout, a regularization technique in deep learning, has shown promise in mitigating overfitting and improving model generalization. This study investigates the impact of different dropout rates on the performance of an ASR system operating on Sepedi- English code-switched data. Through experimentation, varying dropout rates are used to assess the effectiveness of dropout in enhancing the accuracy. The development of the system utilized the Sepedi Prompted Code Switch (SPCS) corpus, and its performance was evaluated with both the National Centre for Human Language Technology Sepedi test corpus and the SPCS corpus. The model generated the lowest Word Error Rate of 45.64%.