Automatic polyphonic piano music transcription by a multi-classification discriminative-learning

In this paper we investigate on the use locally recurrent neural networks (LRNN), trained by a discriminative learning approach, for automatic polyphonic piano music transcription. Due to polyphonic characteristic of the input signal standard discriminative learning (DL) is not adequate and a suitable modification, called multi-classification discriminative learning (MCDL), is introduced. The automatic music transcription architecture presented in the paper is composed by a preprocessing unit which performs a constant Q Fourier transform such that the signal is represented in both time and frequency domain, followed by a peak-peaking and decision blocks: the last built with a LRNN. In order to demonstrate the effectiveness of the proposed MCDL for LRNN several experiments have been carried out.