A Harmonic Structure-Based Neural Network Model for Musical Pitch Detection

Xian Wang, Lingqiao Liu, Qinfeng Shi

Keywords: musical pitch detection, automatic music transcription

TL;DR: harmonic acoustic model

Abstract: In this paper we design a harmonic acoustic model for pitch detection. This model arranges conventional convolution and sparse convolution in a way such that the global harmonic patterns captured by sparse convolution are composed of the enough number of local patterns captured by layers of conventional convolution. When trained on the MAPS dataset, the harmonic model outperforms all existing pitch detection systems trained on the same dataset. Most impressively, when trained on MAPS with simple data augmentation, the harmonic model with an LSTM layer on top surpasses an up-to-date, more complex pitch detection system trained on the MAESTRO dataset to which complicated data augmentation is applied and whose training split is an order-of-magnitude larger than the training split of MAPS. The harmonic model has demonstrated potential to be used for advanced automatic music transcription (AMT) systems.

Code: https://github.com/anyconf/music