There are many proposed studies on the identification of musical instruments, but none been focused on plucked string instruments of the Colombian Andean region such as: tiple, tiple requinto, guitar and bandola.
Therefore, we propose to identify these instruments using machine learning techniques, such as: Linear discriminant analysis (LDA), Decision Tree (DTs), K-nearest neighbor (kNN), Support Vector Machines (SVM), Artificial Neural Networks (ANNs) and using three methods of data reduction: Feature Selection; Principal Component Analysis (PCA) with 10,100 and 1000 major components; and extracting the first five partial frequencies together with their normalized amplitudes.
We carried out this study using our own dataset of 1000 monophonic audio recordings in WAV format of the first position notes of each instrument. Regarding the validation method, we used the Cross Validation with a k equal to five to create the confusion matrices and ROC Curves (Receiver Operating Characteristic). We reached the best values with ANNs that had a 99.8% accuracy percentage, besides the ROC curves showed that the best instrument identified was the guitar.
Download Data set of Musical Instruments: