In this paper, it is assumed that the desired signal f1(t) is a harmonic complex tone, consisting of the fundamental frequency F0(t) and the harmonic components, which are multiplies of F0(t). The proposed model segregates the desired signal from the mixed signal by constraining the temporal differentiation of the instantaneous amplitude, the instantaneous phase, and the fundamental frequency. Here, the relationship between the four regularities[Bregman1993] and the constraints concerned is shown in Table II. These constraints are defined as follows.
Temporal differentiations of the instantaneous amplitude Ak(t), the instantaneous phase
,
and the fundamental frequency F0(t) must be represented by an R-th-order differentiable piecewise polynomial as follows:
F0(t) is the fundamental frequency, and NF0 is the number of harmonics of the highest order.
The harmonic component must satisfy
Suppose that
and
are the onset and offset of the fundamental component.
If the signal component obtained by the k-th channel is the signal component generated by the same acoustic source (that is, harmonic components), then onset
and offset
determined by the k-th channel must coincide with
and
respectively.
That is, the differences in onset and offset must satisfy
Suppose that the amplitude envelope Ak(t) is defined in the closed-duration [ta,tb] and satisfies constraint 1.
If Ak(t) is as smooth as possible, then the following integral must be minimized:
Substituting constraint (9) in Eq. (7), we get the linear differential equation of the instantaneous input phase difference . By solving this linear differential equation, we can determine as follows.
(Proof) See appendix A.
From Lemma 1, if Ck(t) is determined, then is uniquely determined by the above equation. Moreover, if Dk,R(t) is determined, then the two instantaneous input phases can be determined using and Dk,R(t). Therefore, if the two R-th-order polynomials Ck,R(t) and Dk,R(t) are determined as some kind of optimization problem, the two instantaneous amplitudes and the two instantaneous phases can be estimated. Although it is possible to estimate the coefficients Ck,r(t) and Dk,r(t), , there is a problem that the computational cost of estimating two polynomials increases greatly.
In this paper, in order to reduce the computational cost, we assumed that Ck,R(t) is a linear (R=1) polynomial ( dAk(t)/dt=Ck,1(t)) and Dk,R(t) is zero ( ) in constraint 1. In this assumption, the instantaneous amplitude Ak(t) which can be allowed to undergo a temporal change in region, constrains the second order polynomial ( ). Moreover, the instantaneous phase , which is constrained (i.e. ), cannot be allowed to temporarily change. Here, if the number of channels K is very large, each frequency of the signal component that passed through the channel approximately coincides with the center frequency of each channel. Even if the above condition is false, its frequency difference can be represented by Dk,0.
This paper solves the problem of segregating the desired signal f1(t) from the mixed signal, in which noise f2(t) is added to the localized f1(t), under the above assumption.