A finite element model of human middle ear for sound transfer function from tympanic membrane to oval window has been developed and validated and refined using a statistical calibration approach.The middle ear FE model consists of three ossicles (malleus, incus and stapes), tympanic membrane, tendons and ligaments.Two-dimensional images from micro-CT scanning for a human temporal bone were used to build the three-dimensional geometry of the ossicles.The tympanic membrane, ligaments and tendons were then constructed and divided into finite elements.The boundary conditions and material properties of the middle ear FE model have large uncertainty.First, the statistical information on the boundary conditions and material properties were assumed from published data.Next, the variation of stapes velocity transfer function due to the uncertainty of the material properties and boundary conditions were estimated by using the eigenvector dimension method and compared with those of experiments.The comparison shows very good agreement between the two results.However, the measured transfer functions for middle ear show large variation subjects by subjects.To develop a valid FE model for middle ear which can represent the variation, the confidence intervals of the measured transfer functions were extracted and the probability of the stapes velocity transfer function was estimated.Then, a likelihood function between the measured and calculated transfer functions can be defined.By maximizing the likelihood function, the statistical information of the material properties and boundary conditions such as, for example, cochlear fluid spring constant and the stiffness of incudostapedial joint could be identified.Applying the identified material properties and boundary conditions, the finite element model for human middle ear has become to represent the variations of the transfer function as well as expected values.This shows that the proposed validation and refinement procedure is very effective for the human middle ear FE model.