Deep Learning Long Short-Term Memory based Automatic Music Transcription System for Carnatic Music

Abstract

Automatic music transcription uses computational algorithms to covert a music audio into some form of music notation. Automatic music transcription is very important for applications like music mixing, song recommendation and music information retrieval. It is a challenging task involving the domains of signal processing and artificial intelligence. Though there are many works on automatic music transcription for western music, there are very few works for automatic music transcription for Indian classical music, especially Carnatic music. Automatic music transcription for Carnatic music is very challenging due to variations in the swars/note frequencies. Due to the variations, it becomes difficult to detect the Raga and transcribe the music. In this work, a deep learning LSTM based automatic music transcription system is proposed for Carnatic music. The note detection is solved as the image classification problem using a modified Visual Geometry Group Network (VGGNet). The sequence of notes is classified into 72 Melakartha ragas using an LSTM classifier and provides better accuracy when compared to existing methods.