Harmonium note and triad music transcription using neural networks

Learning music requires a two-prong approach which includes theoretical studies and practical exposure to the instrument to be learnt. While previous literature has focused on developing technologies for determining the notes of different musical instruments, the harmonium has not been so popular in this research area. This research focuses on using a hybrid approach for polyphonic triad recognition of the Harmonium music. In this research, over ۲۱۰۰۰ audio samples of harmonium including notes and triads were taken for the Convolutional-Recurrent Neural Network (CRNN) model training purpose. The recorded audio samples were also used to train the Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) models to comparatively analyze the efficiency of these models. The results indicated that the CRNN model is more efficient, accurate, and precise on a score-based transcription. The proposed system produced ۹۴\% accurate results for triad recognition of Harmonium. The recognized triads were represented as sheet music using Lilypond. Possible applications of this output are for better understanding of the triad sequences by students or for Automatic Music Transcription of performances.

Keywords:

Acoustic modeling , Music language modeling (MLM) , Music analysis , Recurrent Neural Networks , Convolutional Neural Network