Metrical Analysis of Musical Audio Using Probabilistic Models

Due to the exploding amount of available music in recent years, media collections cannot be managed manually any more, which makes automatic audio analysis crucial for content-based search, organisation, and processing of data. This thesis focuses on the automatic extraction of a metrical grid, determined by beats, downbeats, and time signature, from a music piece. I propose several algorithms to tackle this problem, all comprising three stages: First, (low-level) features are extracted from the audio signal. Second, an acoustic model transfers these features into probabilities in the music domain. Third, a probabilistic sequence model finds the most probable sequence of labels under the model assumptions. This thesis provides contributions to the second and third stage. I (i) explore acoustic models based on machine learning methods, and (ii) develop models and algorithms for efficient probabilistic inference for both online and offline scenarios. Further, I design applications such as an automatic drummer which listens to and accompanies a musician in a live setting. The most recent algorithms developed in this thesis exhibit state-of-the-art per- formance and clearly demonstrate the superiority of systems incorporating machine learning over hand-designed systems, which were prevalent at the time of starting this thesis. All algorithms developed in this thesis are publicly available as open-source software. I also publish beat and downbeat annotations for the Ballroom dataset to foster further research in this area.