This paper presents a human-voice enhancement method for a deformable and partially-occluded microphone array. Although microphone arrays distributed on the long bodies of hose-shaped rescue robots are crucial for finding victims under collapsed buildings, human voices captured by a microphone array are contaminated by non-stationary actuator and friction noise. Standard blind source separation methods cannot be used because the relative microphone positions change over time and some of them are occasionally shaded by rubble. To solve these problems, we develop a Bayesian model that separates multi-channel amplitude spectrograms into sparse and low-rank components (human voice and noise) without using phase information, which depends on the array layout. The voice level at each microphone is estimated in a time-varying manner for reducing the influence of the shaded microphones. Experiments using a 3-m hose-shaped robot with eight microphones show that our method outperforms conventional methods by the signal-to-noise ratio of 2.7 dB.