Wavelet analysis in musical instrument sound classification
C. Pruysers, J. Schnapp, Ian Kaminskyj · 2006
Given the increased complexity of the non-vibrato recording classifier, in terms of the number of instruments it needs to classify, three different classifier architectures were evaluated: (a) single stage, (b) hierarchic, and (c) hybrid. It is intended that the user of the system will be able to decide which s/he wishes to use, balancing the performance achieved with the computational effort expended. For the purposes of this research, only the single stage classifier was used and the new features were only added to non-vibrato recording classifier since 100% classification accuracy was already easily achievable using the existing vibrato recording classifier features. A study was performed to determine whether the classification accuracy of an existing musical instrument sound classifier could be improved. The existing classifier uses six features to classify monophonic isolated sounds from nineteen different musical instruments. These features summarise important information regarding the spectrotemporal qualities of the musical tones being identified. A classification accuracy of 87.6% was achieved when classifying of 533 non-vibrato recordings from the nineteen instruments. In this research, Morlet wavelets and Dauchencies wavelet packets were added to the existing six features used in the feature extraction and classification stages. Classification of the same test recordings showed that adding the two wavelet features improved the system accuracy by five percentage points. The single stage classifier shown in Figure 2 uses five single feature classifiers to independently determine the most likely instrument to have produced the input test sample. Each classifier extracts its single feature and then employs the kNNC algorithm [2] to classify the input test sound. The single features were used either directly or preprocessed using principal component analysis (PCA).