*4.7. Data Augmentation*

In this section, we test the augmen<sup>t</sup> method mentioned in the former paragraph. The *waveoverlapping* method was used to insert two di fferent crops into a single clip, and then three di fferent kinds of settings were applied. The original ESC50 had 1600 sets of training data in each fold. We randomly picked up certain sets of data to make the additional training clips. Setting 1 took 800 original clips, and Setting 2 took 400 clips. The wave overlapping created two crops from a single original source.

Setting 3 caused these two crops to be pitch shifted or time stretched. We tested the result using a partially pre-trained network and a fully pre-trained one. The results are shown in Tables 9 and 10.


**Table 9.** Test accuracy of augments applied in partial pre-pertained network.

**Table 10.** Test accuracy of augments applied in full pre-trained network.

