Artist filtering for non-western music classification

The "album effect" is a known phenomenon in musical artist and genre recognition. Classification results are often better when songs from the same album are used in the training and evaluation data sets. Supposedly, this effect is caused by the production conditions of the album, e.g. recording quality, mixing and equalization preferences, effects etc. This behavior is not representative of real world scenarios, though, and should therefore be avoided when evaluating the performance of a classifier. The related "artist effect" also affects the results of genre recognition. It is caused by the appearance of the same artists in the training and evaluation data sets. Artist filters have been proposed previously to remove this influence. We perform three different experiments to characterize the "artist effect" somewhat better and to analyze it in conjunction with non-western music. First, we test the effect's influence on the classification of musical pieces into their reg ions of origin. We then repeat this experiment using only specific sets of features (for the timbre, rhythm, and tonality domains). Finally, we perform a finer genre recognition with genres from four different world regions. The influence of the aforementioned effect is evaluated for all experiments.