Analysis of time-frequency representations for musical onset detection with convolutional neural network

In this paper a convolutional neural network is applied to the problem of note onset detection in audio recordings. Two time-frequency representations are analysed, showing the superiority of standard spectrogram over enhanced autocorrelation (EAC) used as the input to the convolutional network. Experimental evaluation is based on a dataset containing 10,939 annotated onsets, with total duration of the audio recordings of over 45 min.