The "winning" system in the 2013 MIREX Latin Genre Classification Task was a deep neural network trained with simple features. An explanation for its winning performance has yet to be found. In previous work, we built similar systems using the BALLROOM music dataset, and found their performances to be greatly affected by slightly changing the tempo of the music of a test recording. In the MIREX task, however, systems are trained and tested using the Latin Music Dataset (LMD), which is 4.5 times larger than BALLROOM, and which does not seem to show as strong a relationship between tempo and label as BALLROOM. In this paper, we reproduce the "winning" deep learning system using LMD, and measure the effects of time dilation on its performance. We find that tempo changes of at most (pm 6,\%) greatly diminish and improve its performance. Interpreted with the low-level nature of the input features, this supports the conclusion that the system is exploiting some low-level absolute time characteristics to reproduce ground truth in LMD.