A Transformer Based Pitch Sequence Autoencoder with MIDI Augmentation

TL;DR

This paper introduces a transformer-based autoencoder with MIDI augmentation for classifying whether a music excerpt is AI-generated or human-composed, achieving competitive results with a compact model.

Contribution

It proposes a novel masked language model approach using ALBERT for composer classification, with data augmentation and refined loss to improve performance on small datasets.

Findings

Achieved 3rd place in CSMT 2020 data challenge

Reduced model parameters while maintaining accuracy

Demonstrated effectiveness of data augmentation in small datasets

Abstract

Despite recent achievements of deep learning automatic music generation algorithms, few approaches have been proposed to evaluate whether a single-track music excerpt is composed by automatons or Homo sapiens. To tackle this problem, we apply a masked language model based on ALBERT for composers classification. The aim is to obtain a model that can suggest the probability a MIDI clip might be composed condition on the auto-generation hypothesis, and which is trained with only AI-composed single-track MIDI. In this paper, the amount of parameters is reduced, two methods on data augmentation are proposed as well as a refined loss function to prevent overfitting. The experiment results show our model ranks $3^{rd}$ in all the $7$ teams in the data challenge in CSMT(2020). Furthermore, this inspiring method could be spread to other music information retrieval tasks that are based on a small dataset.