This is the result of overfitting the data with a bootsrapping method. 4X the data size training set was extracted by randomly sampling the training data with replacement.

Single layer MGU, 16 note sequences bootstrapX4 data, 1.4°

This album features an window into the Drum-Off outputs and the corresponding player drum parts transcribed by the game. Clear changes to model output can be heard when at part *24.wav the player starts playing a 16th note hi-hat pattern and at *28.wav where the player switches to a different beat again. The model used in this experiment is a "Parallel Time Distributed Convolution MGU with Attention Layers" model. The model was initialized at game start and the running number before ".wav" is the number of training examples the model has received. The Orig* files are the game transcribed player parts and the Gen* files are the model generated comebacks to those performances. Temperature was set to 0.95° for this experiment.

Orig ATT TDC P Mgu 29

Gen 0.75° ATT TDC P Mgu 28

Gen 0.75° ATT TDC P Mgu 27

Gen 0.75° ATT TDC P Mgu 26

Gen 0.75° ATT TDC P Mgu 25

Gen 0.75° ATT TDC P Mgu 24

Gen 0.75° ATT TDC P Mgu 23

Gen 0.75° ATT TDC P Mgu 22

Drum-Off transcription of a part played with a loud acoustic kit.

Gen Temp Orig 39

New multi input - multi output RNN generation. This model is very capable of producing drumbeats.

Gen 0.75° Multi2multi 39

Computer generated drum tracks. The A.I. listens to a live drum playing and learns different drum patterns by separating different sources from mono audio. These patterns are then fed to a neural network that learns the structure of drumbeats and outputs new midi tracks upon request. The work is highly incomplete and all the new tracks offer a status report on the program. At the time of writing the computer is working with 4 minutes of live drum audio, when the input system is working the amount of data will increase as I, The programmer, get around playing a bit more.

NMFD->MGU->Midi

Popular tracks