Abstract
Within music information retrieval (MIR) research, numerous beat‑tracking systems have been developed, targeting either audio recordings or symbolic representations such as MIDI files. However, the differences between these approaches, their respective strengths and weaknesses, and the potential for combining them have received limited attention. In this article, we compare two conceptually different beat trackers: an audio‑based model that operates frame by frame and a symbolic‑based model using an event‑driven approach. Specifically, we analyze the performance of two pretrained systems: the audio beat tracker madmom and the symbolic beat tracker Performance MIDI‑to‑Score (PM2S). Our evaluation is based on a cross‑modal dataset of Chopin's Mazurkas (Maz‑5), which includes multiple audio recordings and MIDI representations automatically transcribed from audio. As a key contribution, we standardize the post‑processing pipelines for the frame‑based and event‑based beat trackers to ensure comparability and explore various late‑fusion methods within a unifying framework. Our results highlight the effectiveness of these fusion strategies in leveraging the strengths of both modalities while providing valuable insights into the performance of existing beat‑tracking models.