Robustness of Features and Classification Models on Degraded Data Sets in Music Classification

There exists a large number of supervised music classification tasks: Recognition of music genres and emotions, playing instruments, harmonic and melodic properties, temporal and rhythmic characteristics, etc. In recent years, many studies were published in that field, which are either focused on complex feature engineering or application and tuning of classification algorithms. How- ever, less work is done on the evaluation of model robustness, and music data sets are often limited to music with some common characteristics, so that the question about the generalisation ability of proposed models usually remains unanswered. In this study, we examine and compare the classification perfor- mance of audio features and classification models when applied for recognition of genres and instruments on music data sets which were degraded by means of techniques available in the Audio Degradation Toolbox including attenuation, compression, live and vinyl recording degradations, and addition of noise.