In the context of MIREX 2005 contest, our lab have participated to 4 competitions out of 10 and got the first position on two of them. Here I'm going to present the note onset detection one. The goal consists of finding the time location of the maximum number of notes onsets events, while minimizing the number of spurious onsets. To accomplish this task, we use machine learning to classify musical frames into onset or non-onset. The strength of this technique is to provide a robust algorithm to a wide variety of music styles. This characteristic makes it suitable for many other tasks with almost no tuning. It can be used for tempo extraction, music finger printing, music editing ... In the presentation, I'll first presents the details of the algorithm and the different techniques tried. Then I'll presents the results of the contests and some tests performed on a separated custom dataset.