Contribution
Perceptual Drift in Speech Quality: Reassessing Listener Judgments a Decade Later
* Presenting author
Abstract:
With advancements in telecommunication technology, including efficient codecs and shifts from narrowband to full-band transmission, modern listeners—accustomed to richer audio fidelity—may rate older speech samples more critically, suggesting an evolution in quality expectations. This study examines perceptual drift over time in listener evaluations of speech quality by re-assessing a corpus of speech samples originally rated over a decade ago. Using the same audio files and a listener demographic similar to the original study, this work examines changes in perceived overall quality to assess shifts in listener standards over time. Ratings are collected on an identical scale and statistically compared with the original scores to determine if significant differences emerge.Such an investigation holds important implications for machine learning models trained on historical subjective scores, which serve as ground truth in speech quality assessment. Modern evaluation systems, such as NISQA and ITU-T’s emerging P.SAMD, are trained on datasets that include speech samples from as far back as the mid-2000s. Evidence of perceptual drift could indicate a need to recalibrate or retrain these models to align with evolving standards. Understanding these potential shifts ensures that subjective testing methods and corresponding models remain relevant, supporting accurate evaluations in future speech quality research.