Automatic cover detection -- the task of finding in a audio dataset all covers of a query track -- has long been a challenging theoretical problem in MIR community. It also became a practical need for music composers societies requiring to detect automatically if an audio excerpt embeds musical content belonging to their catalog. In a recent work, we addressed this problem with a convolutional neural network mapping each track's dominant melody to an embedding vector, and trained to minimize cover pairs distance in the embeddings space, while maximizing it for non-covers. We showed in particular that training this model with enough works having five or more covers yields state-of-the-art results. This however does not reflect the realistic use case, where music catalogs typically contain works with zero or at most one or two covers. We thus introduce here a new test set incorporating these constraints, and propose two contributions to improve our model's accuracy under these stricter conditions: we replace dominant melody with multi-pitch representation as input data, and describe a novel prototypical triplet loss designed to improve covers clustering. We show that these changes improve results significantly for two concrete use cases, large dataset lookup and live songs identification.

Automatic cover detection -- the task of finding in an audio database all the covers of one or several query tracks -- has long been seen as a challenging theoretical problem in the MIR community and as an acute practical problem for authors and composers societies. Original algorithms proposed for this task have proven their accuracy on small datasets, but are unable to scale up to modern real-life audio corpora. On the other hand, faster approaches designed to process thousands of pairwise comparisons resulted in lower accuracy, making them unsuitable for practical use. In this work, we propose a neural network architecture that is trained to represent each track as a single embedding vector. The computation burden is therefore left to the embedding extraction -- that can be conducted offline and stored, while the pairwise comparison task reduces to a simple Euclidean distance computation. We further propose to extract each track's embedding out of its dominant melody representation, obtained by another neural network trained for this task. We then show that this architecture improves state-of-the-art accuracy both on small and large datasets, and is able to scale to query databases of thousands of tracks in a few seconds.