RNN Pop Music Generation

Problem

Symbolic music generation usually tokenizes absolute pitch, which means the model has to learn the same melodic shape separately in every key. That inflates the vocabulary, splits the training signal across redundant tokens and wastes capacity on a transformation that is musically trivial.

Data

• Source: POP909, 909 pop songs in MIDI, converted through an automated tokenization pipeline
• Representation: interval-based tokens: the distance between notes rather than absolute pitch
• Output format: generated tokens are piped back to sheet music via music21, so results are immediately readable and playable

Approach

A recurrent network over interval tokens. Instead of emitting the pitch $p_t$ directly, the sequence is re-expressed as the difference between consecutive notes:

$$\Delta_t = p_t - p_{t-1}$$

Encoding transitions rather than positions makes the representation key-invariant: a phrase transposed into a different key is the same token sequence, so every training example teaches the model about every key at once. That change alone reduced vocabulary size by 80%.

The pipeline is end-to-end automated: MIDI in, packaged training sequences out, generated sheet music at the other end.

Experiments

What didn’t work

Artifacts

• RepoSource on GitHub →
• DatasetPOP909 dataset →