We propose to carry out comprehensive analysis and testing of the algorithms abstracted from the above models. A main focus of the proposed work is to exploit desirable properties of the auditory spectrum that remain unutilized in acoustic detection and recognition applications such as those discussed in Thrust Area V. To achieve this goal, we shall reformulate the auditory models into efficient workable signal processing algorithms accessible to the engineering community, and compatible with as much of the existing infrastructure as possible [.Shamma Lyon 1996, Lazzaro 1996, Teolis 1993, Shamma Wang normalize 1994.]
As an example of this proposed approach, we shall consider a particular
auditory algorithm for estimating the acoustic spectrum [.Shamma Wang
normalize 1994.]. This algorithm is based on a minimal model of
the most important early auditory stages. These consist of an affine
wavelet transform (W) of a pre-emphasized sound signal (
). This signal (
) is acted upon by a compressive
sigmoidal nonlinearity (
), and a derivative with respect to
the scale axis of the transform (
). The intermediate output
thus generated is given by:
This signal is then half-wave rectified and integrated to produce the final auditory spectrum:
where 
is a temporal integration window [.Shamma Wang normalize
1994, Yang Shamma Wang 1992.]. This auditory spectrum turns out to be
effectively a self-normalized acoustic spectrum with significant
spectral peak enhancements and robustness against scaling and noise
corruption.
As part of our Thrust Area V projects, we plan to demonstrate these advantages in a variety of applications ranging from the identification of targets with battlefield acoustic sensor arrays, to machine monitoring and diagnostics. We are confident, based on preliminary investigations [.Shamma Wang normalize 1994, Shamma Byrne Robinson 1989.] that the auditory spectrum will surpass other acoustic representations in preserving target signal features down to lower signal to noise ratios, and hence facilitate superior detection, identification, and tracking.
We shall also explore the viability of real-time implementations using both software and hardware realizations (see Thrust area IV for details). This is an important prerequisite for the wide acceptance and dissemination of the algorithms. Some non-real-time software realizations are now available in well documented and efficient code that can be readily accessed [.Ru 1996, Lyon Slaney 1988.].