Using ERPs to Assess the Processing of Words Under Broadcast Bit Rate Limitations

Abstract

In telecommunication research, audio quality is typically assessed with behavioral tests. Neurophysiological data can complement these as an objective and non-intrusive measure, potentially revealing neuronal differences in quality processing below the threshold of conscious perception that might affect a user’s long-term satisfaction. Recently, subconscious processing of noise in phonemes was found in event-related potentials (ERPs). The present EEG study (N = 8) applies this approach to a more realistic setting. In a forced-choice task, subjects had to rate whether a given word was of maximal quality or degraded. Stimuli were presented either in wideband quality (60%) or were impaired by four progressive levels of bit rate reduction using a standard telecommunication codec. The ERP analysis allowed qualifying those trials where subjects correctly indicated a loss of quality: the higher “neural uncertainty” involved in detecting more subtle degradations is reflected in a decreased amplitude of the P3 component. Additionally, a linear classifier was used to identify trials in which quality impairment was not reported but still evoked an ERP pattern similar to conscious detection (in three subjects). This suggests that even when no quality impairment is consciously noted, the degradation can still be processed in the cortex. The results show that ERP paradigms developed for noisy phonemes can be successfully extended to full words degraded by realistic broadcast limitations. This supports the potential of ERP analysis combined with machine learning to not only complement conventional audio quality assessment methods but also extend their sensitivity to sub-threshold stimuli.