TABLE I.

Calculation of the estimated target spectrogram in different two-stage networks. ${\mathcal{G}}^{(1)}(\cdot )$ and ${\mathcal{G}}^{(2)}(\cdot )$ denote the first and second stage DNN. ${\mathit{\mu }}_{O_1}$ and ${\mathit{\sigma }}_{O_1}$ denote normalization parameters for the output of the first stage, and ${\mathit{\mu }}_{O_1}$ and ${\mathit{\sigma }}_{O_2}$ the parameters for the second stage.

Combination	DNN formula
Mapping+Mapping	$\left|{\widehat{S}}_{\text{1}}\right|=exp\left({\mathcal{G}}^{(2)}\left(\left[\frac{{\mathcal{G}}^{(1)}(\overline{\mathit{F}}(m))-{\mathit{\mu }}_{o_1}}{{\mathit{\sigma }}_{o_1}},\overline{\mathit{F}}(m)\right]\right)\times {\mathit{\sigma }}_{o_2}+{\mathit{\mu }}_{o_2}\right)$
Mapping+Masking	$\left|{\widehat{S}}_{\text{1}}\right|={\mathcal{G}}^{(2)}\left(\left[\frac{{\mathcal{G}}^{(1)}(\overline{\mathit{F}}(m))-{\mathit{\mu }}_{o_1}}{{\mathit{\sigma }}_{o_1}},\overline{F}(m)\right]\right)\times \left|Y\right|$
Masking+Mapping	$\left|{\widehat{S}}_{\text{1}}\right|=exp\left({\mathcal{G}}^{(2)}\left(\left[\frac{log\left({\mathcal{G}}^{(1)}(\overline{\mathit{F}}(m))\times \left|Y\right|\right)-{\mathit{\mu }}_{o1}}{{\mathit{\sigma }}_{o_1}},\overline{\mathit{F}}(m)\right]\right)\times {\mathit{\sigma }}_{o_2}+{\mathit{\mu }}_{o_2}\right)$
Masking+Masking	$\left|{\widehat{S}}_{\text{1}}\right|={\mathcal{G}}^{(2)}\left(\left[\frac{log\left({\mathcal{G}}^{(1)}(\overline{\mathit{F}}(m))\times \left|Y\right|\right)-{\mathit{\mu }}_{o_1}}{{\mathit{\sigma }}_{o_1}},\overline{\mathit{F}}(m)\right]\right)\times \left|Y\right|$