A Novel Approach to String Instrument Recognition

In music information retrieval, identifying instruments has always been a challenging aspect for researchers. The proposed approach offers a simple and novel approach with highly accurate results in identifying instruments belonging to the same class, the string family in particular. The method aims to achieve this objective in an efficient manner, without the inclusion of any complex computations. The feature set developed using frequency and wavelet domain analyses has been employed using different prevalent classification algorithms ranging from the primitive k-NN to the recent Random Forest method. The results are extremely encouraging in all the cases. The best results include achieving an accuracy of 89.85% by SVM and 100% accuracy by Random Forest method for four and three instruments respectively. The major contribution of this work is the achievement of a very high level of accuracy of identification from among the same class of instruments, which has not been reported in existing works. Other significant contributions include the construction of only six features which is a major factor in bringing down the data requirements. The ultimate benefit is a substantial reduction of computational complexity as compared to existing approaches.