Self-Supervised Mean Opinion Score Prediction of Phase-Vocoder-Based Virtual Bass System

The virtual bass system (VBS) leverages a psychoacoustic phenomenon known as the “missing fundamental” to trick listeners into perceiving the fundamental frequency from its higher harmonics. The VBS finds common use in consumer electronic devices where miniature and flat panel loudspeakers are integrated, as they cannot reproduce satisfactory low-frequency components. The additional harmonics introduced by the VBS can lead to perceptual distortion. Therefore, evaluating the perceptual quality of the VBS necessitates subjective listening tests. Previous studies have attempted to derive objective metrics and identify combinations of model output variables to predict the perceptual quality of the VBS. However, due to the limited number of subjective test results used to obtain the combination coefficients, inconsistencies may arise in predictions. This paper proposes to adopt self-supervised deep learning models to predict the mean opinion score (MOS) of the VBS. Experiment results demonstrate a strong linear correlation between the model outputs and the human-rated MOS, indicating that a linear mapping is sufficient to convert a model output into an accurate MOS prediction.