The inner ear labyrinth is a combined sensory organ of hearing and balance, which is surrounding the bony cavity located in the petrous temporal bone. The structure of the inner ear labyrinth plays an important role in otology research and clinic diagnosis of ear diseases. Automatic and accurate segmentation of the inner ear labyrinth is a foundation of computer-aided temporal bone quantitively measurements and diagnosis. The inner ear labyrinth is characterized by its complex morphology, small size, and high labeling cost, which brings challenges for deep learning-based automatic segmentation methods. In this paper, we propose a robust segmentation method for the labyrinth in temporal bone CT images via multi-model inconsistency. In the active learning paradigm, we design an informative sample assessment strategy for screening informative unlabeled data. An observer network is introduced to confirm the confidence of segmented voxels based on the inconsistency to a backbone segmentation network. To further improve the efficiency of the sample screening, a maximum-connected probability map (MCP-Map) is introduced to eliminate the influence of outliers in the result of coarse segmentation. Experimental results show that our methods have the highest labeling efficiency and the lowest labeling cost compared with several existing active learning methods. With 40% labeled reduce, our method achieved 95.67% in Dice Similarity Coefficient (DSC), which is the state-of-the-art in the labyrinth segmentation.