Tone detection for Indian classical polyphonic instrumental audio using DNN model

Abstract

Identification of tone from a polyphonic audio is quite a challenging task in digital audio processing. When the audio clip is a classical instrumental track the process is even more cumbersome. This paper proposes a novel approach to detect the tone of polyphonic Indian classical instrumental audio using scaled exponential linear unit (SeLu) activated Deep Neural Network (DNN) along with instrument identification which also uses SeLu activated DNN Model. This aims at utilising the same key features which help in instrument detection in real-life situations. The number of features were also reduced from 34 to 26 in comparison with the earlier work by analysing and identifying the redundant features and adding a few more important characteristic features. The proposed Instrument identification model predicts instruments with an accuracy of 84.39% for Carnatic classical and 83.59% for Hindustani classical. The SeLu activated DNN model for tone detection has attained an accuracy of 88.30%.