Application of Elman Back Propagation Neural Network for Automatic Identification of Tabla Strokes in North Indian Classical Music

Tabla is the most useful accompanying percussion instrument used in North Indian Classical Music. The homophonic sound produced from Tabla instruments generates multiple harmonics. Therefore, Tabla stroke identification is a challenging task. Tabla stroke identification has various applications such as Tempo Estimation, Beat Tracking, Rhythm Identification, Tala Recognition, Automatic Music Transcription to name a few. This research aims to compare Elman Neural Network (ENN) with various neural network architectures useful for automatic Tabla stroke identification. This comparison would be useful to appropriately select the ENN for the applications of Tabla stroke identification. Studio recordings of 640 Tabla audio excerpts, sampled at 44, 100 Hz sampling frequency are used to train and test the neural networks namely, Feed Forward Back Propagation Neural Network (FFBPNN), Pattern Recognition Neural Network (PRNN), Elman Neural Network (ENN), Cascade Forward Neural Network (CFNN), and Recurrent Neural Network (RNN). The audio features are extracted using traditional Mel Frequency Cepstral Coefficient (MFCC) and Timbral audio descriptors along with MFCC. The Tabla strokes are categorized into two major categories namely, Open and Closed Tabla strokes. The result shows that Tabla stroke identification accuracy is obtained higher for open strokes due to the difference of Attack, Decay, Sustain, and Release values of the strokes. The Tabla stroke identification accuracy of 94.1% is achieved using ENN, for Timbral audio features.