Accordion detection in folk music recordings

In the thesis, we implemented an algorithm that automatically recognizes an instrument in Slovene Folk music, in our case the accordion. The input argument of our algorithm is an arbitrarily long wave form recording. Its 3-seconds long sections are then classified into one of two groups – contains the accordion or does not contain the accordion. We have built the learning base for machine learning on the basis of database containing field recordings of Slovene Folk music. We cut 4680 3-seconds long sections from long recordings, 2340 of those incorporated the accordion and 2340 did not. We decided to devote the first third of the base to learning, the second third of the base to testing and the last third of the base to calculating the relevance of the characteristics of sound. We have calculated the characteristics of sound in all recordings and then researched which features of sound affect the classification the most. We have neglected redundant features od sound and thus increased the speed and accuracy of classification. We have taught the classifier with the method of support vectors and the help of learning database and then tested it with our test database of recordings. The testing showed the classifier to be 95,83% accurate. At the end, we implemented the algorithm in MATLAB environment