Timbre Based Source Separation and Identifica...

Timbre Based Source Separation and Identification of Tabla Strokes using Multiple Neural Networks

Author Name : Shambhavi Shete, Saurabh Deshmukh

ABSTRACT Music Information Retrieval (MIR) has gained popularity after the recent advancements in digital signal processing techniques and machine learning. Today, there exist many automatic music analysis techniques. Singer identification, Song recognition, Musical instrument identification, Chord recognition, and Music transcription are some of the important applications of MIR. Tabla is one of the most popular percussion instruments used in North Indian Classical Music. The acoustic instrument Tabla is a combination of two drums. The strokes are produced from either Left, Right or Both the drums simultaneously, producing a homophonic texture of the sound. Timbre is a non-tangible fourth dimension of sound that uniquely defines it. In this research, we have implemented two Neural Network systems. A Multiple Neural Network system that uses the Timbral attributes of the sound to separate the Tabla strokes based on their sources and Identifies the Tabla stroke and a Single Neural Network that classifies the Tabla stroke into nine classes are compared here. The former system exhibits a better performance to identify the Tabla stroke over the later. When compared to the performance of both these systems in terms of ac- curacy, the Multiple Neural Network system gives an accuracy of 91.1% as compared with the Single Neural Network system that gives an accuracy of 82.2%.