I trained WaveGAN neural network on 7500 vintage drumloops (breaks) from around 2000 old recordings. All the loops were fit to 120bpm tempo and sampled at 32768Hz for the dataset. The trained model is able to generate either non-existing breaks or new drumloops with elements of existing breaks, and smoothly morph between generated loops. This result is about to be improved in the future. Project by: @ With help from: @ Colab and model: This sequence in wav: (you may sample this if you want, but I'm not 100% sure that all the samples are copyright-free or smth) The video features a sequence of 4400 breaks played at 3 different speeds: 0:00:00 Fast Breakbeat tempo (147 bpm) 1:59:54 Slow Breakbeat tempo (120 bpm, original pitch) 4:26:41 Jungle/DnB tempo (160 bpm) (diluted with speech samples
Hide player controls
Hide resume playing