cnmat/MIDI_tokenization_models

1---2license: mit3tags:4- music-generation5- midi6- transformer7- symbolic-music8datasets:9- maestro10- asap11---12 13# Evaluating Tokenization Strategies for Expressive Classical Piano Performance Generation14 15This 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)*.16 17We provide six composer/genre-conditional MIDI Transformers, each trained on a distinct tokenization scheme to systematically evaluate expressive piano performance generation.18 19## Model Variants & Tokenization Modes20 21| Subfolder | Tokenization Mode | Target Features |22| :--- | :--- | :--- |23| `note` | `note` | Onset, Pitch & duration |24| `note_pedal` | `note_pedal` | Onset, Pitch, duration, sustain pedal |25| `note_velocity` | `note_velocity` | Onset, Pitch, duration, velocity |26| `note_velocity_beat` | `note_velocity_beat` | Onset, Pitch, duration, velocity, beat markers |27| `note_velocity_pedal` | `note_velocity_pedal` | Onset, Pitch, duration, velocity, sustain pedal |28| `full` | `full` | Complete expressive performance attributes |29 30## Training Strategy31 32* **Pre-training:** **Lakh MIDI** + **MAESTRO** dataset  33  *(Note: Lakh MIDI was incorporated during pre-training to improve generation diversity, expanding upon the baseline setup described in the paper).*34* **Fine-tuning:** **ASAP** (Aligned Scores and Performances) dataset.35