This thesis investigates the application of transfer learning and embedding-based approaches to music genre classification, addressing the challenge of limited labeled data in music information retrieval. We explore three complementary approaches using the GTZANdataset: a baseline multilayer perceptron with hand-crafted audio features, a convolutional neural network leveraging VGGish embeddings pre-trained on YouTube audio, and a k-nearest neighbors classifier operating in the embedding space. Analysis of confusion patterns provides insights into genre boundaries and overlaps, suggesting that the embedding space effectively captures musical similarity beyond rigid genre categorization. We conclude by proposing a framework for transforming the genre classifier into a music recommendation system by utilizing the learned embeddings for similarity-based retrieval, potentially enabling more nuanced music discovery that transcends traditional genre limitations.