This paper presents a text independent speaker authentication method adapted to mobile devices. Special attention was placed on delivering a fully operational application, which admits a sufficient reliability level and an efficient functioning. To this end, we have excluded the need for any network communication. Hence, we opted for the completion of both the training and the identification processes directly on the mobile device through the extraction of linear prediction cepstral coefficients and the naive Bayes algorithm as the classifier. Furthermore, the authentication decision is enhanced to overcome misidentification through access privileges that the user should attribute to each application beforehand. To evaluate the proposed authentication system, eleven participants were involved in the experiment, conducted in quiet and noisy environments. Public speech corpora were also employed to compare this implementation to existing methods. Results were efficient regarding mobile resources’ consumption. The overall classification performance obtained was accurate with a small number of samples. Then, it appeared that our authentication system might be used as a first security layer, but also as part of a multilayer authentication, or as a fall-back mechanism.