This work proposes novel stability-preserving model order reduction approaches for vibro-acoustic finite element models. As most research in the past for these systems has focused on noise attenuation in the frequency-domain, stability-preserving properties were of low priority. However, as the interest for timedomain auralization and (model based) active noise control increases, stability-preserving model order reduction techniques are becoming indispensable. The original finite element models for vibro-acoustic simulation are already well established but require too much computational load for these applications. This work therefore proposes two new global approaches for the generation of stable reduced-order models. Based on proven conditions for stability preservation under one-sided projection, a reformulation of the displacement-fluid velocity potential (u ) formulation is proposed. In contrast to the regular formulation, the proposed approach leads to a new asymmetric structure for the system matrices which is proven to preserve stability under one-sided projection. The second approach starts from a displacement-pressure (u p) description where the system level projection space is decoupled for the two domains, for which we also prove the preservation of stability. Two numerical validation cases are presented which demonstrate the inadequacy of straightforward model order reduction on typical vibro-acoustic models for time-domain simulation and compare the performance of the proposed approaches. Both proposed approaches effectively preserve the stability of the original system. formulation is that it results in a real dynamic stiffness matrix for undamped frequency-domain simulations and in a real eigenvalue problem for the computation of the undamped modes. It is the standard formulation used in most commercial FE tools for coupled vibro-acoustic analysis [10].One of the major drawbacks of element-based methods is their computational cost, which scales unfavourably with increasing problem size and frequency, restricting the overall practical applicability. In order to accurately describe the wave-like spatial character of the variables, the use of locally defined and low-order shape functions imposes minimum requirements on the number of elements used per wavelength and therefore limits the maximum element size for a given wavelength [11]. The highest frequency of interest determines the smallest wavelength present in the model, which in turn dictates the maximum element size. Because of the three-dimensional nature of acoustics, the number of elements in the model grows rapidly with increasing problem size or equivalently with decreasing element size (and thus also with increasing frequency). In order to alleviate the problems associated with computational costs, one can turn to model order reduction (MOR) techniques to reduce the number of DOFs for a desired accuracy [12]. For coupled vibro-acoustic problems, the advantages of MOR have already been demonstrated in frequency-domain analysis [10,13,14]. I...