Maximum a posteriori estimation of spectral gain with harmonic-structure-based phase reconstruction for phase-aware speech enhancement.