Transformer-Based Multimodal Framework for Music Similarity Analysis and Recommendation Systems

Modern music streaming platforms offer vast catalogs and personalized discovery experiences. Nevertheless, current music recommendation systems often overemphasize popular content and fail to capture complex user preferences or support the exploration of niche genres. This paper addresses these limitations by proposing a deep learning–based multimodal recommendation framework that leverages transformer architectures to analyze audio signals and contextual metadata. The goal is to enhance music similarity modeling and recommendation accuracy by generating enriched embeddings that capture musical structure, instrumentation, and genre subtleties. The research introduces a system that combines audio-based features with metadata through a fusion strategy informed by attention mechanisms. The methodology includes large-scale experimentation on public music datasets and evaluation using standard recommendation quality metrics. Results demonstrate improved personalization and diversity in recommendations compared to baseline models. This work contributes to the field by providing a novel multimodal architecture and demonstrating the effectiveness of audio-content-aware recommendation strategies.