Computational sound quality models are trained on subjective listening data
Evidence
The problem with subjective evaluation is that the required experiments are typically very laborious and time consuming. The experiments should be carried out in conditions that correspond to real usage of the sounds or related products. Thus, it would be much easier and less expensive to use objective criteria or models for sound quality. Ideally, a computational model provides an estimate of quality from signal analysis which correlates well with subjective data. The development of such a model (Figure 17.1) is typically based on proper subjective listening experiments from which data are collected and analysed statistically. Using various techniques, the factors and features of sounds that best explain the subjective behaviour are sought. The computational model can be a simple linear regression model or a more complex non-linear model, such as a neural network trained to map the feature parameters to quality indices. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2332
Categorization
- XREF: Connects to the book’s overall theme of computational auditory models (e.g., auditory periphery modeling earlier in Pulkki) predicting perceptual outcomes. — connects to: Computational auditory models serve knowledge, audio-quality measurement, and diagnostics — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 821; Computational simulation can generate testable hypotheses about auditory physiology; Predicting subjective listening preference from measured data remains a hard unsolved problem. (+5 more)
- corroborated: 2 independent passages support this claim — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2332
Broader Connections
- Auditory-model difference measures predict audible quality degradation
#corroborates(0.85) - Objective sound quality models have limited domains of validity and cannot replace subjective testing
#qualifies(0.80) — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2337 - Percentile loudness N5 predicts how listeners judge overall noise immission loudness
#exemplifies(0.75) — Professor Dr.-Ing. Hugo Fastl, Professor Dr.-Ing. Eberhard, Psychoacoustics- Facts and Models, 1799