Table of Contents
Advertisement
C H A P T E R
5 :
W I D E
B A N D
A U D I O
Q U A L I T Y
T E S T I N G
the Advanced version is hidden in the respective ear models and the set of
MOVs used. Both versions comprise an
artificial neural network
for the
cognitive modelling. Since these networks are usually critical in terms of
reliability, special care was taken not to over-train the network during the design
phase. Subsequent investigations proved the stability and plausibility of the
networks.
5.4.3 Basic version
The "Basic" version implements an FFT based ear model, as outlined in Figure
5.3.
Most features of this model are based on the fundamental psychoacoustic
principles as described by Zwicker [ZWIC67, ZWIC82]. After the functional
verification of the model, the developers tried to fine tune various parameters,
but contrary to other existing proposals, the optimum was found very close to
the theoretical values. Other models to-date always had to compensate for
simplifications of the implementation by a misadjustment of other parameters.
Figure 5.3: Perceptual model, PEAQ "Basic"
Following the signal flow from the input signal to the final calculation of the
excitation pattern, the processing starts by a transformation of the input signal to
the frequency domain. A 2048 point FFT is applied along with subsequent
scaling of the spectra, according to the listening level, which has to be input by
the user as a parameter. This results in a frequency resolution of approximately
23.4 Hz, and a corresponding temporal resolution of 23.4 ms (at 48 kHz
sample rate).
In the consecutive block the effects of the outer and middle ear are modelled by
weighting the spectrum with the appropriate filter functions. Afterwards the
spectra are grouped into critical bands, achieving a resolution of 1/4 bark per
band. The subsequent adding of "internal noise" is intended to model effects,
such as the permanent masking of sounds in our auditory system caused by the
streaming of blood and other physiological phenomena. This step is followed
by the calculation of masking effects. Simultaneous masking is modelled by a
frequency and level dependent spreading function according to [TERH79] with
slight modifications. Temporal masking is modelled only partly since the
88
Advertisement
Table of Contents
