A Dataset for Multimodal Music Information Retrieval of Sotho-Tswana Music Videos

The dataset is a multimodal dataset for music information retrieval of under-resourced Sotho-Tswana language. The MIR tasks for which the dataset can be used for, include multimodal music sentiment analysis, multimodal music genre classification, multimodal language identification in music videos, and recommender systems. Although the primary source of the musical videos is YouTube, several processes have been performed on the dataset, such as segmentation, separation of the audio modality from the visual modality, and extraction of spectral-based acoustic features, together with enriched manual annotations of the following metadata, language, genre, lyrics and meaning of song. The dataset is ideal for training deep learning models, and the recommended fusion method for combining the features of the audio and visual modalities is the decision-level (late) fusion method. In addition to the musical video clips, part of the dataset includes various CSV files and Jupyter notebooks.