Static loudness models fail on time-variant sounds, motivating frame-based temporal integration models

Evidence

Section 10.2 introduced a simple loudness model that accurately estimates the loudness perception evoked by many relatively simple signals. Such models have been found to be unable to derive accurate loudness estimates for spectrally complex and time-variant signals. This shortcoming has motivated the design of more advanced loudness models (Florentine et al., 2005), of which two approaches are briefly touched upon next. The first one (Zwicker, 1977) processes the signal in the time domain and produces a continuous signal representing the loudness as a function of time. This approach opens up the possibility of predicting the loudness perception evoked by a time-variant signal based on the peak values in the model output. The second model (Glasberg and Moore, 2002) divides the signal into overlapping time frames and extracts short-term loudness values from each time frame. The model also emulates the temporal integration of loudness between adjacent time frames, and consequently, the model provides accurate estimates for time-variant signals as well, although this estimate is derived by simply averaging across the short-term loudness values. Despite the improved accuracy in predicting perceived loudness of complex signals, the different models cannot yet fully explain loudness perception. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 1659

Categorization

Category: Causal Relationships

• QUESTION: What exactly does Glasberg-Moore’s ‘temporal integration’ emulate physiologically — neural adaptation, persistence? Worth chasing before citing.

Related

• Two competing loudness models predict time-variant loudness by peak tracking or frame averaging (similarity 0.94)
• Perceived loudness of time-varying sounds tracks near-peak levels, not averages (similarity 0.86)
• Loudness models need 100-200 ms time constants to capture temporal integration and post-masking. (similarity 0.86)
• Long-term averaging obscures transient attack and release sounds (similarity 0.85)
• Sound level meters trade average level against tracking speed via time response. (similarity 0.85)
• Perceived loudness of time-varying sounds is set by short-time peak values. (similarity 0.84)
• Time-varying sound loudness is set by integrated loudness peaks, not moment-to-moment level (similarity 0.84)
• A simple high-pass filter can replace elaborate loudness models (similarity 0.83)
• Loudness perception takes time to develop and so can be interrupted by later sounds (similarity 0.82)
• Different hearing tasks exploit different model time constants. (similarity 0.80)
• Steady-state loudness models fail for speech, music, and impulsive technical noises. (similarity 0.80)
• Hearing forms loudness by integrating sound energy over about 100-200 ms (similarity 0.79)
• Short sounds decay in perceived loudness faster than long sounds. (similarity 0.79)
• Temporal integration models divide on time constant assumptions of hundreds versus single-digit milliseconds. (similarity 0.79)
• Loudness processing acts as a temporal low-pass filter with about 45 ms integration time (similarity 0.79)