cnmat/MIDI_tokenization_models

Evaluating Tokenization Strategies for Expressive Classical Piano Performance Generation

This repository contains official artifacts for the paper "Evaluating Tokenization Strategies for Expressive Classical Piano Performance Generation", submitted to the 29th International Conference on Digital Audio Effects (DAFx26).

We provide six composer/genre-conditional MIDI Transformers, each trained on a distinct tokenization scheme to systematically evaluate expressive piano performance generation.

Model Variants & Tokenization Modes

Subfolder	Tokenization Mode	Target Features
note	note	Onset, Pitch & duration
note_pedal	note_pedal	Onset, Pitch, duration, sustain pedal
note_velocity	note_velocity	Onset, Pitch, duration, velocity
note_velocity_beat	note_velocity_beat	Onset, Pitch, duration, velocity, beat markers
note_velocity_pedal	note_velocity_pedal	Onset, Pitch, duration, velocity, sustain pedal
full	full	Complete expressive performance attributes

Training Strategy

• Pre-training: Lakh MIDI + MAESTRO dataset
  (Note: Lakh MIDI was incorporated during pre-training to improve generation diversity, expanding upon the baseline setup described in the paper).
• Fine-tuning: ASAP (Aligned Scores and Performances) dataset.