Audio Beat Tracking
An Analysis of Beat Tracker Accuracy in Audio Streaming Scenarios
AbstractThis study looked at the effects of audio streams on the accuracy of several beat trackers. Past research on beat tracking typically tested beat trackers on individual audio clips in isolation. This doesn’t evaluate beat tracker accuracy on streams of concatenated audio clips. This study took the publicly available Essentia Multifeature Beat Tracker, INESC-Porto Beat Tracker, Madmom Beat Tracker, Madmom Beat Detector, DBN Beat Tracker, and CRF Beat Detector and evaluated their accuracy in streaming and non-streaming scenarios. The Ballroom, GTZAN, and SMC audio datasets were used as a source of beat annotated audio files for both streaming and non-streaming tests. Comparing accuracy in streaming and non-streaming scenarios showed that all beat trackers had a lower accuracy on audio streams than on individual audio clips. The INESC-Porto Beat Tracker performed the worst in both scenarios, whilst the DBN Beat Tracker performed the best. The Madmom Beat Detector and CRF Beat Detector performed well in the non-streaming scenarios but their accuracy dropped significantly in the streaming scenario. Looking at beat tracker accuracy over time showed the INESC-Porto Beat Tracker and Madmom Beat Detector steadily lost accuracy over time during audio streams, whilst the other beat trackers maintained roughly constant accuracy. Comparing the DBN Beat Tracker to the Madmom Beat Tracker showed that the improved Dynamic Bayesian Network of the DBN Beat Tracker improved its accuracy in both streaming and non-streaming scenarios. Comparing the CRF Beat Detector to the Madmom Beat Detector showed that the CRF Beat Detector’s probabilistic post processing step improved accuracy significantly in streaming scenarios.
How to Cite
STRUTT, Ben. Audio Beat Tracking. Discovery, Invention & Application, [S.l.], june 2017. Available at: <https://computing.derby.ac.uk/ojs/index.php/da/article/view/256>. Date accessed: 24 june 2019.