African Beats and Lyrics Composition Using Recurrent Neural Networks and Transformer-Based Models

Abstract

Human music composition frequently struggles with challenges such as ensuring originality, maintaining lyrical coherence, and managing elaborate polyrhythms. These difficulties are particularly evident in African music, which is characterized by complex rhythms, call and response structures, rhythmic improvisation, and tonal language. This study explores how artificial intelligence can support and enhance African music composition by integrating Recurrent Neural Networks (RNNs) and transformer based models. The hybrid system was designed to generate both beats and lyrics, drawing on pre trained models such as Mistral 7B and MusicGen. The methodology adopted includes model pre training, data preparation, adaptation, composition pipelining, inference generation, post processing, and evaluation. Results demonstrated that the hybrid model achieved improved rhythmic consistency, lyrical coherence, and Afrobeat authenticity. Quantitative evaluation indicated lower error rates and stronger stylistic regularity, while subjective listener assessments averaged 4.29/5, indicating positive reception. When compared with existing models, the hybrid approach balanced structural coherence with creative diversity, offering a culturally grounded framework for AI driven African music composition.