Beat Tracking for Salsa Music: Adapting and Benchmarking Models Using a Newly Introduced Salsa Dataset

This study addresses the challenge of adapting current beat tracking algorithms, predominantly trained on Western music, to the rhythmic complexities of Salsa, a genre rich in syncopations and polyrhythms. Using training meth- ods that minimise the need for extensive annotated data, we benchmark the adaptability of two established models: BeatNet and BöckTCN, on our newly introduced beat and downbeat annotated Salsa dataset. We find that, on Salsa music, models trained with Salsa largely outperform mod- els trained without any Salsa, nearly matching the accuracy of these models on Western music. This research not only establishes a baseline for beat and downbeat tracking per- formance in Salsa music but also contributes to the broader goal of developing more adept music information retrieval systems. We also contribute a 40-song Salsa dataset for beat and downbeat tracking research in this genre.