A COMPARISON OF PERCEPTUAL RATINGS AND COMPUTED AUDIO FEATURES
Publication Type:Conference Paper
Source:SMC Conference 2011 (2011)
The backbone of most music information retrieval sys-tems is the features extracted from audio. There is an abundance of features suggested in previous studies rang-ing from low-level spectral properties to high-level se-mantic descriptions. These features often attempt to model different perceptual aspects. However, few studies have verified if the extracted features correspond to the assumed perceptual concepts. To investigate this we se-lected a set of features (or musical factors) from previous psychology studies. Subjects rated nine features and two emotion scales using a set of ringtone examples. Related audio features were extracted using existing toolboxes and compared with the perceptual ratings. The results indicate that there was a high agreement among the judges for most of the perceptual scales. The emotion ratings energy and valence could be well estimated by the perceptual features using multiple regression with adj. R2 = 0.93 and 0.87, respectively. The corresponding audio features could only to a certain degree predict the corre-sponding perceptual features indicating a need for further development.