Automatic Piano Music Transcription by Hadamard Product of Low-Rank NMF and CNN/CDAE Outputs

抄録

In this paper, we propose novel automatic piano music transcription methods which improve multi-pitch estimation accuracy. The conventional multi-pitch estimation method using low-rank non-negative matrix factorization (LR-NMF) cannot achieve a good accuracy because the method cannot detect true spectrum and analyze nonlinearity in sound. In the proposed method, the nonlinearity is analyzed using the convolutional neural network (CNN) or the convolutional denoising autoencoder (CDAE) as a post processing of LR-NMF. After the processing, we further improve the accuracy by the Hadamard product of the output from LR-NMF and that from CNN or CDAE. The performance of the proposed method is evaluated through computer simulation.