Context-Independent Polyphonic Piano Onset Transcription with an Infinite Training Dataset

TL;DR

This paper introduces a data synthesis method for polyphonic piano onset transcription that enables training neural networks without large real datasets, improving generalization across recording conditions.

Contribution

It presents a novel data generation approach that models piano dynamics and avoids dataset limitations, enhancing transcription performance and generalization.

Findings

Achieves good transcription accuracy on MAPS dataset

Demonstrates excellent generalization to new recordings

Avoids dataset curation and disentanglement issues

Abstract

Many of the recent approaches to polyphonic piano note onset transcription require training a machine learning model on a large piano database. However, such approaches are limited by dataset availability; additional training data is difficult to produce, and proposed systems often perform poorly on novel recording conditions. We propose a method to quickly synthesize arbitrary quantities of training data, avoiding the need for curating large datasets. Various aspects of piano note dynamics - including nonlinearity of note signatures with velocity, different articulations, temporal clustering of onsets, and nonlinear note partial interference - are modeled to match the characteristics of real pianos. Our method also avoids the disentanglement problem, a recently noted issue affecting machine-learning based approaches. We train a feed-forward neural network with two hidden layers on our generated training data and achieve both good transcription performance on the large MAPS piano dataset and excellent generalization qualities.