Overview of the proposed model

First, the observed signal f(t) is decomposed into S_k(t) and $\phi_k(t)$ using an auditory-motivated filterbank. This filterbank is implemented as a constant Q gammatone filterbank, constructed with K=128, bandwidth of 60-6000 Hz, and sampling frequency of 20 kHz [8]. Next, the fundamental frequency F₀(t) of the desired signal is determined using an amplitude spectrogram S_k(t)s (see Sec. 3.1). Then, the concurrent time-frequency region of the desired signal is determined using constraints (i) and (iii) [8]. In the determined concurrent time-frequency region, A_k(t) and B_k(t) are determined from S_k(t), $\phi_k(t)$, $\theta_{1k}(t)$, and $\theta_{2k}(t)$. S_k(t) and $\phi_k(t)$ are determined by using the amplitude and phase spectra defined by the wavelet transform [8]. $\theta_{1k}(t)$ and $\theta_{2k}(t)$ are determined using constraints (ii) and (iv) (see Sec. 2.2 and 3.2). Finally, f₁(t) and f₂(t) are determined from Eqs. () and (), respectively.

 

Table: Constraints corresponding to Bregman's psychoacoustical heuristic regularities.

Regularity	Constraint
(i) common onset/offset	synchronous of onset/offset
(ii) gradualness of change	piecewise-differentiable
	polynomial approximation
(slowness)	(Kalman filtering)
(smoothness)	(spline interpolation)
(iii) harmonicity	multiples of the
	fundamental frequency
(iv) changes occurring in	correlation between the
the acoustic event	instantaneous amplitudes