Generating music using a multi-layer LSTM

Our (hypothetical) creative agency client loves what we've done in how we can generate music lyrics. Now, they want us to create some music. We will be using multiple layers of LSTMs, as shown in the following diagram:

By now, we know that RNNs are good for sequential data, and we can also represent a music track as notes and chord sequences. In this paradigm, notes become data objects containing octave, offset, and pitch information. Chords become data container objects holding information for the combination of notes played at one time.

Pitch is the sound frequency of a note. Musicians represent notes with letter designations [A, B, C, D, E, F, G], with G being the lowest and A being the highest.

Octave identifies the set of pitches used at any one time while playing an instrument.

Offset identifies the location of a note in...