Most of the video has them using an algorithm to predict the sound of the hits, then finding the closest match in the database & using it as a sample. At 1:50,

Most of the video has them using an algorithm to predict the sound of the hits, then finding the closest match in the database & using it as a sample. At 1:50, they switch to using the prediction as a synthesised replacement. It is not as good. They switch back to the first method for the tests