Music Generator

• Data:
  1. .mp3, analysing soundwaves
  2. .mid, analysing scores
• Processing: Mixture Density Network (MDN), Density function approximation
• Output: autoregressive music predictor

WAV soundwaves

To process soundwaves in a standard mp3 file you will have to worry about 44k datapoints per second. Even by reducing quality that is very muchas muchachas for learning time.

For that I took a Neural Network with bayesian interpretation of the output units.

Sinus

The model can solve toyproblems, like generating Sinus.

drum

The model can solve toyproblems, like imitate a drum.

original

generated

rap

The model can solve toyproblems, like extracting the baseline of a music piece.

original

generated

Musical Instrument Digital Interface (MIDI)

Patterns are always hierarchical. Thinking about the environment, I don’t want to predict in a complexity of a soundwave. Lets go MIDI.

By reducing the dimensions of the environment the model is capable of predicting a MIDI file in perfect manner. The Mixture Density Network creates a Random distribution over the Music peace. What you basically hear in the following are random samples drawn from the probability distribution of the given song. The easiest example is a coinflip. Given a large enough sequence of coinflips my mixture density network would approximate the probability distribution and you would have a random generator for a coinflip. Exactly the same is happening for the music peaces. Just in a higher dimensional manner.

Starwars

My favorite (from 1.5min+). So much different from the original.

original

generated

Scale A major

Simple music basics for the agent to learn with interesting result. Maybe one has to give the agent music patterns and no music pieces…

original

generated

Tetris

Very exciting, it is archived as the first working descision process.

original

generated