Embedded speaker recognition in mobile devices involves a limited amount of computing resource but also is linked with several ergonomic constraints. For example both the enrolment and the test have to be done using short audio sequences. Even if they proved their efficiency in more classical situations, GMM/UBM based systems show their limits in this context. This paper deals with this problem and proposes to take into account the linguistic nature of the speech material inside the GMM/UBM framework. The proposed solution mixes the text-independent aspects of the GMM/UBM with a semi-continuous like approach in order to deal with the text-dependent information. This system respects both the resource and the ergonomic constraints of the considered application field. The preliminary experiments are done on the publicly available database ValidDB and show the potential of the proposed approach. Particularly, when compared to the GMM/UBM, our approach decreases drastically both the computational cost and the equal error rates when impostors don't know the user passwords. For other situations the performance remains comparable between both approaches.