Investigating relationship between sound quality metrics and spectro-temporal modulation spectrogram

Sound quality metrics (SQMs), such as loudness, sharpness, roughness, and fluctuation strength, are used to estimate sensory pleasantest and psychoacoustical annoyance. As these metrics are influenced by both temporal and spectral characteristics of audio signals, a framework capturing joint time-frequency features is essential. This study investigates relationship between SQMs and spectro-temporal modulation (STM) spectrogram, which represents patterns across time and frequency. We used a dataset of 2000 environmental sounds from ESC-50, computed STM spectrograms, and estimated SQMs using auditory-based computational models. Partial least squares regression was applied to identify STM spectrogram regions most relevant to each SQM. Bootstrapping was used to derive 95% confidence intervals for variable importance in projection scores. The analysis revealed distinct STM spectrogram contributions for each SQM: (1) Loudness and sharpness corresponded to static components near 0 Hz temporal and spectral modulation, (2) roughness was mainly affected by temporal modulations around 70 Hz, and (3) fluctuation strength was associated with low-frequency modulations near 4 Hz. These findings suggest STM spectrogram provides a unified framework for evaluating and potentially manipulating perceptual aspects of sound quality.