This project develops a content-based music recommender that predicts song relationships from audio, using listener-generated tags as supervision during training. From 10-second clips, stem separation and mel spectrograms are used to represent each track, and a late-fusion ResNet18 learns embeddings that capture genre, mood, and musical structure. At inference time, the system recommends songs from audio alone through an interactive web app, showing how deep learning can support music discovery without relying solely on user behavior.