Speech-to-text and text-to-speech conversions are referred to as Automatic Speech Recognition (ASR), which employs transcripts to transform speech signals into text. Convolutional Neural Networks (CNNs) have considerable efficacy in ASR, particularly in the feature extraction stage, where they outperform traditional models in capturing local hierarchical features from spectrogram representations of audio signals. Nonetheless, the efficacy of standard CNN across multiple thresholds in speech-to-text conversion for low-resource languages remains under investigation. This study examines the effects of two distinct thresholds (0.22 and 0.35) in a CNN-based model for Yorùbá speech-to-text conversion. A corpus of 64 unique Yorùbá language samples was utilized for implementation and evaluation in MATLAB R2023a. The model is evaluated using False Positive Rate (FPR), Specificity (Spec), Sensitivity (Sen), Precision (Prec), Accuracy (Acc), and F1-score. The FPR diminishes by 0.19%, but the SPEC ascends by 0.19%, signifying enhanced true negative detection at the elevated threshold. The speech-to-text conversion of the Yorùbá language dataset was precise, with a reduction in SEN of 0.12%. The PREC climbs by 0.22% as the threshold increases, indicating an enhancement in the model's accuracy for affirmative case predictions. An increase of 0.5% in the Acc signifies an enhancement in overall performance. A more balanced performance is evidenced by a 0.05% enhancement in the F1 Score, which equilibrates precision and recall. Performance metrics reveal a modest enhancement of 0.35. The results underscore the capacity of language technology to embody global linguistic diversity and enhance the accessibility of speech-to-text conversion systems.