Proposal and Evaluation of a Pitch Estimation Model Based on Reconstruction of the Spectrum of Input Sound by Using Timbre Information and Pitch Information

Abstract

Automatic music transcription (AMT) is a task that automatically generates music scores from the sound of music. Pitch estimation is one of the most important technologies in AMT, and its accuracy strongly affects the quality of AMT. Although its accuracy is improving, thanks to deep neural networks nowadays, pitch estimation for multi-instrumental music is still a challenging task, because there are many difficulties caused by numerous and complicated timbre patterns in multi-instrumental music. In this paper, we propose a pitch estimation model based on the reconstruction of the spectrum of input sound which can reduce the effect of timbre complexity on pitch estimation. The experimental result indicated that the proposed model improved the average precision by 8.8 points compared to the conventional method.