A Statistic Learning Approach to Tempo Estimation for Audio Music

Automatic beat tracking and tempo estimation are challenging tasks, especially for audio music with nonbinary tempo or weak percussion. This paper proposes a Kmeans clustering approach to handle tempo estimation with one-third/triple tempo or weak percussion. In particular, the first stage is to compute the tempo curve from the tempogram by DP(Dynamic Programming). Then use Kmean clustering to extract tempo candidates with 2 to 4 clusters. Finally, make tempo rules discovery to match ground truth of audio dataset by learning approach. The tempo estimation algorithm could almost obtain maximum 1.0 pscore value for mirex2006 tempo training dataset with 20 excerpts in length of 30 seconds. Index Terms – Tempo Estimation, Tempogram , Tempo Curve, K-means Clustering, Dynamic Programming