Audio Onset Detection Using Machine Learning Techniques: the Effect and Applicability of Key and Tempo Information

This paper explores the effect musical context, namely key and tempo, on audio onset detection using machine learning techniques, with a focus on the changes in performance caused by mismatched key and tempo between training and test pieces, and the potential benefits of incorporating such musical information. We extract frequency energy information from audio as the input instance attributes for the machine learning techniques. The audio is synthesized from MIDI files, which provide exact onset times. We test two state-of-the-art machine learning algorithms, Support Vector Machines and Neural Networks, for their learning and classification of audio onsets. In the first experiment, testing is performed on the training piece, transposed to different keys and time-stretched to various tempi. The results show that audio onset detection performs significantly better when the key and tempo of the test and training sets concur, than when they are different. We propose several ways to incorporate key and tempo in an onset detection system. In the second experiment, we use J. S. Bach’s 24 Preludes from the Well-Tempered Clavier Book 1. The inclusion of tempo information improves the results on average. The positive performance changes supports the usefulness of tempo and key knowledge in the design of onset detection systems, or in the prescription of confidence statistics to onset detection outcomes.