Loudness of time-varying environmental sounds: Still a challenge for current loudness models?
Invited paper
Otto von Guericke University Magdeburg
Wednesday 3 june, 2015, 10:00 - 10:20
0.6 Madrid (49)
Abstract:
Environmental sounds are often non-stationary, i.e., they vary in their
physical properties over time. A prominent example is speech but also
various technical sounds have such a non-stationary character. The present
study focusses on the loudness of these sounds by comparing own measured
loudness data with predictions of existing loudness models.
Dynamic loudness models commonly assume that the instantaneous loudness is
smoothed with one or more time constants and that the overall loudness is
determined by the maximum or a certain percentile of this smoothed loudness
time function. The time constant and the measures used to derive overall
loudness differ between the different loudness models. For example, the
model of Chalupper and Fastl (2002) uses a 8-Hz low-pass filter to derive a
short-term loudness time function. The model of time-varying loudness by
Glasberg and Moore (2002) even proposes two loudness measures within the
same model, the short-term and long-term loudness, where the latter is
determined by longer time constants than the former. This long-term loudness
was specifically developed to predict the loudness of amplitude modulated
sounds. It is, however, unclear which measure is more appropriate to predict
loudness of environmental sounds. Levels at equal loudness between reference
sounds and speech as well as technical sounds were measured and compared to
the predicted overall loudness of the two models mentioned above. The data
indicate that loudness of speech-like signals seems to be largely determined
by a loudness function that was smoothed with longer time constants than
currently used in dynamic models, since the long-term spectrum largely
determined the loudness. The loudness of time-varying, technical signals
could not be fully predicted with any of the existing models but again
longer time-constants appear to be beneficial for the prediction. However,
discrepancies remain.