Abstract
Mixing music is the process of combining tracks of recorded audio to an overall piece. This is a complicated process and, hence, automatic mixing or metering tools would be useful. The aim of the current research project was to work towards measuring the perceived quality of music mixes by establishing predictors for one important perceptual attribute of high- quality mixes (spectral clarity). A review of academic and non-academic literature revealed that the high-level parameters that are responsible for determining the perceived quality of a music mix are ‘clarity and separation’, ‘balance’, ‘impact and interest’ and ‘freedom from technical faults’, alongside context-specific parameters. A further in-depth literature review established that clarity and separation—the chosen focus for this research—depend on spectral, spatial and intensity factors, and temporal changes in these factors. Spectral factors play an important role across all areas of literature consulted (namely timbral clarity, clarity in concert halls, masking, loudness, auditory scene analysis and speech intelligibility), and so the impact of mix EQ on spectral clarity was investigated in a series of experiments. These experiments determined that two important factors contribute to the spectral clarity of single sounds. These are the harmonic centroid (spectral centroid divided by the sound’s average fundamental frequency) and mid-range spectral peakiness (related to sharp peaks in the frequency spectrum). For sounds modified by simple spectral filtering, these two factors are sufficient to model clarity changes with a Spearman correlation ranging from 0.631 (bass and vocal stimuli) to 0.848 (string stimuli). For sounds in a mix, however, other factors become important. Adding a peak audibility measure proved useful. This measure determined whether the audibility of peaks in the spectra of the target sounds was increased or decreased through EQ. Target and overall mix harmonic centroids and mid- range spectral peakiness, combined with peak audibility, correlated positively with target spectral clarity (r=0.568). Findings could contribute to the development of marketable products such as a piece of software able to judge the overall sound quality of a mix, automatic mixers or sonically improved music production software. Further work will allow a more comprehensive and generalizable model to be developed.