In this paper, automatic musical instrument identification using a variety of classifiers is addressed. Experiments are performed on a large set of recordings that stem from 20 instrument classes. Several features from general audio data classification applications as well as MPEG-7 descriptors are measured for 1000 recordings. Branch-and-bound feature selection is applied in order to select the most discriminating features for instrument classification. The first classifier is based on non-negative matrix factorization (NMF) techniques, where training is performed for each audio class individually. A novel NMF testing method is proposed, where each recording is projected onto several training matrices, which have been Gram-Schmidt orthogonalized. Several NMF variants are utilized besides the standard NMF method, such as the local NMF and the sparse NMF. In addition, 3-layered multilayer perceptrons, normalized Gaussian radial basis function networks, and support vector machines employing a polynomial kernel have also been tested as classifiers. The classification accuracy is high, ranging between 88.7% to 95.3%, outperforming the state-of-the-art techniques tested in the aforementioned experiment.