Brian Strope
bps@ucla.edu
A First-Order Model of Dynamic Auditory Perception
Brian Strope and Abeer Alwan
Department of Electrical Engineering, UCLA
A computational model of dynamic audition is essential to understand the
perception of non-stationary speech. A series of level- and
frequency-dependent forward-masking perceptual experiments, using
sinusoidal maskers and probes, and an adaptive "transformed up-down"
2AFC paradigm, provide data to quantify model parameters. Masker levels
range between 40 and 90 dB SPL, masker to probe delays are from 15 to
120 ms, and stimuli frequencies are from 250 to 4000 Hz. The model
(similar in structure to a multi-band compression hearing aid) consists
of a filter bank combined with carefully-parameterized logarithmic
automatic gain control on each frequency band to model adaptation in
human audition. The error of the model prediction of forward masking
thresholds has a standard deviation of 2.5 to 3.3 dB across frequencies.
The adaptive model provides a tool to analyze and predict short-term
dynamic, or context-dependent, auditory perception. Such a model has
important implications for the dynamic characteristics of non-linear
hearing aids.
A postscript version of
the poster
is available (or through:
ftp
).
home page
bps@ucla.edu