Transcription of Arabic and Turkish Music Using Convolutional Neural Networks

Automatic music transcription is considered a complex operation, especially in polyphonic music tracks, in addition. However, most of the previous studies hasn't taken into account quarter notes detection which is considered the most important feature in Arabic and Turkish music. In this paper, a complete polyphonic pitch tracking system has been built for Arabic, Turkish, and Western music which is able to detect notes and framewise from piano music tracks. This model combines both signal processing and deep learning techniques to provide more accurate results. In the first stage, constant Q Transformation is applied to the input musical track and then passed to the next stage which consists of two convolutional neural networks (CNNs). The first network is designed to detect notes onset and the other network is for framewise pitch detection. In the end, the output of the neural network is processed to generate MIDI file for the input Music track.