Segregating Musical Chords for Automatic Music Transcription: A LSTM-RNN Approach

Abstract

Notating or transcribing a music piece is very important for musicians. It not only helps them to communicate among each other but also helps in understanding a piece. This is very much essential for improvisations and performances. This makes automatic music transcription systems extremely important. Every music piece can be broadly categorized into two parts namely the lead section and the accompaniment section or background music (BGM). The BGM is very important in a piece as it sets the mood and makes a piece complete. Thus it is very much important to notate the BGM for properly understanding and performing a piece. One of the key components of BGM is known as chord which is constituted of two or more musical notes. Every composition is accompanied with a chord chart. In this paper, a long short term memory-recurrent neural network (LSTM-RNN)- based approach is presented for segregating musical chords from clips of short durations which can aid in automatic transcription. Experiments were performed on over 46800 clips and a highest accuracy of 99.91% has been obtained for the proposed system.