Detecting musical versions (different renditions of the same piece) is a challenging task with important applications. Because of the ground truth nature, existing approaches match musical versions at the track level (e.g., whole song). However, most applications require to match them at the segment level (e.g., 20s chunks). In addition, existing approaches resort to classification and triplet losses, disregarding more recent losses that could bring meaningful improvements. In this paper, we propose a method to learn from weakly annotated segments, together with a contrastive loss variant that outperforms well-studied alternatives. The former is based on pairwise segment distance reductions, while the latter modifies an existing loss following decoupling, hyper-parameter, and geometric considerations. With these two elements, we do not only achieve state-of-the-art results in the standard track-level evaluation, but we also obtain a breakthrough performance in a segment-level evaluation. We believe that, due to the generality of the challenges addressed here, the proposed methods may find utility in domains beyond audio or musical version matching.

Navigating large audio collections presents a significant challenge due to the intricate nature of sound properties and the varied needs of users. To enhance user experience, audio-sharing platforms offer a sound similarity function, which leverages vector-based representations of audio clips to facilitate the retrieval of sounds. This study evaluates the retrieval performances of one manually engineered audio representation and thirteen deep audio embeddings in the semantic sound similarity task. By employing a diverse range of models, our research investigates the effects of utilizing different input modalities and training objectives. In the process, we explore various design choices for integrating embeddings into sound similarity systems. Our evaluation is based on objective ranked performance metrics that incorporate sound classes and sound families, complemented by preliminary subjective assessments. We observe that the multimodal models using audio and language modalities outperform audio-only models by a significant margin, which in turn outperform audio and image models. Notably, the state-of-the-art models on the sound event classification task are not the topperforming models on the semantic sound similarity task. In addition, our findings in embedding processing methods and similarity search functions provide insights broadly applicable to information retrieval systems across different modalities.