Listening Level Changes Music Similarity

We examine the effect of listening level, i.e. the absolute sound pressure level at which sounds are reproduced, on music similarity, and in particular, on playlist generation. Current methods commonly use similarity metrics based on Mel-frequency cepstral coefficients (MFCCs), which are derived from the objective frequency spectrum of a sound. We follow this approach, but use the level-dependent auditory spectrum, evaluated using the loudness models of Glasberg and Moore, at three listening levels, to produce auditory spectrum cepstral coefficients (ASCCs). The ASCCs are used to generate sets of playlists at each listening level, using a typical method, and these playlists were found to differ greatly. From this we conclude that music recommendation systems could be made more perceptually relevant if listening level information were included. We discuss the findings in relation to other fields within MIR where inclusion of listening level might also be of benefit.