Investigating CNN-based Instrument Family Recognition for Western Classical Music Recordings

Western classical music comprises a rich repertoire composed for different ensembles. Often, these ensembles consist of instruments from one or two of the families woodwinds, brass, piano, vocals, and strings. In this paper, we consider the task of automatically recognizing instrument families from music recordings. As one main contribution, we investigate the influence of data normalization, pre-processing, and augmentation techniques on the generalization capability of the models. We report on experiments using three datasets of monotimbral recordings covering different levels oft imbralcomplexity: isolated notes, isolated melodies, and polyphonic pieces. While data augmentation and the normalization of spectral patches turned outtobebeneficial, pre-processing strategies such as logarithmic compression and channel-energy normalization did not lead to substantial improvements. Furthermore, our cross-dataset experiments indicate the necessity of further optimization routines such as domain adaptation.