Music recognition tools have been one of the primary research problems in music information retrieval. Recent advancements in neural networks have shown that unsupervised and semi-supervised learning frameworks can be used to learn latent embeddings of musical data which, in turn, can be used as compact and unique fin-gerprints for query search and retrieval. These frameworks are data-driven, scalable to real-world applications, and robust to noisy environments and transformations which may, advertently or inadvertently, make its way into the query audio record-ings in a real-world setting. This work implements a self-supervised contrastive learning architecture for audio fingerprinting which can scalable to a large music dataset. Motivated by the domain-knowledge of music, this work also proposes modifications which would improve its performance in particularly two aspects: ro-bustness to pitch-shifting and time-stretching.