This study investigates the effect of reverberation on the accuracy of a musical instrument recognition model based on Line Spectral Frequencies and K-means clustering. 180 experiments were conducted by varying the type of music databases (isolated notes, solo performances), the stage in which the reverberation is added (learning and/or testing), and the reverberation type (3 different Reverberation Times, 10 different dry-wet levels). The performances of the model systematically decreased when reverberation was added at the testing stage (by up to 40%). Conversely, when reverberation was added at the training stage, a 3%-increase of performance was observed for the solo performances database. The results suggest that pre-processing the signals with a dereverberation algorithm before classification may be a means to improve musical instrument recognition systems.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.