Beat and Downbeat Tracking Based on Rhythmic Patterns

Example 1: Audio, Annotations, Pattern, Ellis, Dixon, IBT, Klapuri

This first example shows one of the typical behaviours of general-purpose beat-tracking algorithms when dealing with Candombe. The chico drum has an invariable pattern with a strong accent on the second tatum beat. Therefore, this is sometimes predicted as the location of the tactus beat, as can be seen in the figure, in which the predictions are approximately aligned around the second tatum beat, except for the proposed algorithm.

Example 2: Audio, Annotations, Pattern, Ellis, Dixon, IBT, Klapuri

There is another characteristic strong accent performed by the piano drum which is located at the fourth tatum beat. In this second example, this accented stroke is the first strong audio event at the beginning of the plot. Then, it can also be identified in the next measure, four tactus beats ahead. As can be seen, some algorithms tend to predict the beat at this point. In addition, there are also some errors in detecting the correct beat period.

Example 3: Audio, Annotations, Pattern, Pattern adapted

This is an example in which the proposed algorithm does not correctly track the beat. This may indicate that the informed rhythmic pattern to track does not properly match this performance. To check this hypothesis, a K-means clustering is carried out only with this recording. Then, the centroid of the majority cluster is used as the informed pattern to track. The performance of the proposed algorithm is notably improved in this way as can be seen in the figure (labeled as Adapted).