<p>RIYE Audio Classification: A 1D-CNN Baseline for Ogun State Yoruba Dialects</p>

Introduction Dialect classification remains a challenging task in low-resource language settings, particularly where annotated audio data are scarce. The Yoruba language, widely spoken in Nigeria, exhibits significant dialectal variation across regions, including Remo, Ijebu, Yewa, and Egba (RIYE) in Ogun State. Despite their linguistic similarities, these dialects differ in phonetic and intonational patterns, making automated classification both relevant and complex. Methods This preliminary study proposes a one-dimensional convolutional neural network (1D-CNN) baseline approach for classifying RIYE dialects using expert-curated audio data. Due to the limited size of the original dataset, data augmentation techniques, including pitch shifting, time stretching, and temporal segmentation, were applied to programmatically expand the dataset into a variation-rich cohort. Audio signals were transformed into frequency-domain representations using the Short-Time Fourier Transform (STFT) to capture dialect-specific features. The 1D-CNN was trained on a balanced subset of the augmented dataset, and model performance was evaluated using accuracy, precision, recall, and Area Under the Curve (AUC). Results The augmented dataset significantly improved class balance, enabling effective model training. The proposed 1D-CNN baseline model achieved strong performance, with training metrics of 98.79% accuracy, 99.92% AUC, 0.0407 loss, 98.86% precision, and 98.79% recall, and validation metrics of 98.33% accuracy, 99.91% AUC, 0.0586 loss, 98.41% precision, and 98.17% recall. Furthermore, the findings indicate that frequency-domain representations derived from STFT, when combined with convolutional architectures, are implicitly highly effective for capturing subtle phonetic variations and enabling accurate dialect classification in low-resource settings. Conclusion The findings demonstrate the effectiveness of deep learning techniques for dialect classification in low-resource environments, achieving high baseline performance despite limited initial data availability. The integration of data augmentation and frequency-domain feature extraction proved critical in enabling robust learning. However, the reliance on augmented data underscores the need for larger, high-quality real-world datasets to enhance generalizability and mitigate potential synthetic bias. This study provides a strong foundation for future research in Yoruba dialect classification and other under-resourced linguistic domains.