PEAQ uses a trained neural-network cognitive model to predict human MOS ratings.

Evidence

The perceptual evaluation of audio quality (PEAQ) method (Thiede et al., 2000) is an evolved version of the PAQM computation. PEAQ includes an auditory model implemented either with the computationally lighter DFT processing or with the temporally more accurate filter-bank processing. The computation to estimate the MOS has two stages. First, an auditory model is used to compute auditory features, such as excitation patterns, specific loudness patterns, and modulation patterns, for each auditory frequency band. The features are computed for both the reference signal and the signal reproduced using a device and the simulated or real network. The estimated features and the metrics describing differences in the features between the reference and test cases are provided as an input to a cognitive model, which then estimates the MOS value. The cognitive model is, in principle, a pattern recognition algorithm, such as an artificial neural network. The algorithm is trained using a large set of examples from real devices and acoustics conditions which have been analysed by a large panel of listeners. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2371

Categorization

• XREF: Resonates with how trained pattern-recognition substitutes for human panels, paralleling trust/reputation systems that formalize subjective judgments.
• QUESTION: How does the PEAQ training data (real devices, listener panels) get collected, and what ceiling does it hit vs. blind listening?