Overview of the proposed model

An overview of signal flow in the proposed model is shown in Fig. 3.

First, this model can observe the mixed signal f(t) (Fig. 3. A). The observed signal is decomposed into an instantaneous amplitude S_k(t) and an instantaneous phase $\phi _k(t)$ using an auditory-motivated filterbank (Fig. 3. B, C). Next, the fundamental frequency F₀(t) of the desired signal is determined using an amplitude spectrogram (or an other method) (Fig. 3. D). Next, the concurrent time-frequency region of the segregation target is determined using grouping constraints. The frequency region, in which harmonic exist, is determined using the fundamental frequency F₀(t) and the constraint of harmonicity (Fig. 3. E, and focus on a-a') . The time region in which harmonics exist is determined using the constraint of common onset and offset (Fig. 3 F, and focus on b-b').

Next, in the determined concurrent time-frequency region, the instantaneous amplitude A_k(t) and $\theta_{1k}(t)$ of the desired signal f₁(t) are determined from the instantaneous amplitude S_k(t) and the instantaneous output phase $\phi _k(t)$. These are determined by estimating C_k,1(t) and D_k,0, which are constraining A_k(t) and $\theta_{1k}(t)$, respectively (Fig. 3. H, I). In particular, C_k,1(t) is estimated using the constraints of smoothness and correlation between the instantaneous amplitudes A_k(t). And D_k,0 is estimated using the estimated C_k,1(t) and constraint of correlation between the instantaneous amplitudes A_k(t).

Finally, the segregated signal is reconstructed the grouping the instantaneous amplitude A_k(t) and the instantaneous input phase $\theta_{1k}(t)$ (Fig. 3. J).