Spatialized audio stimuli must preserve separate spectra per channel to estimate binaural local SNR

Evidence

In order to investigate the contributions of target speech at various levels of local SNR, the target speech needed to be manipulated. The first step in the stimulus manipulation consisted of the calculation of the local, i.e. spectro-temporal, SNR of the three-speaker stimulus. Spectrograms were calculated from 1024-point fast Fourier transforms on 1024-samples long, 50 %-overlapping, square root Hann-windowed segments of the signals. This was performed for each of the two stereo channels of a spatialized speech signal to accommodate interaural level differences. This resulted in a pair of left and right spectrograms for each signal, from which one estimate of the spectro-temporal power distribution needed to be calculated. This was achieved by computing the squared mean of the absolute values of each set of two (i.e. left and right) corresponding Fourier components and repeating this for the complete spectro-temporal plane (Fig. — Dijk, Pim van, Physiology, Psychoacoustics and…, 1636

Categorization

• QUESTION: Worth understanding whether the squared-mean-of-channels approach to collapsing stereo spectra is the standard choice, or whether more sophisticated binaural SNR metrics exist.