Ever-growing systems' complexity and novel requirements engineering approaches such as reuse or globalization imply that requirements are produced by different stakeholders and written in possibly different languages. In this context, checking consistency so that requirements specifications are amenable to formal analysis is a challenge. Current techniques either fail to consider the requirement set as a whole, missing certain inconsistency types or are unable to take heterogeneous (i.e. expressed in different languages) specifications into account. We propose to use model composition to address this problem in a staged approach. First, heterogeneous requirements are translated in model fragments instances of a common metamodel. Then, fragments are merged in one unique model. On such a model inconsistencies such as under-specifications can be incrementally detected and formal analysis is made possible. Our approach is fully supported by our model composition framework. We propose model composition as means to address flexibility needs in requirements integration. Threats to validity such as the impact of new requirements languages needs to be addressed in future work.