This paper describes the implementation of the IRISA unit selection-based TTS system for our participation in the Blizzard Challenge 2015. For our first participation, we chose to take part in the first task by building one voice for each of the 5 Indian languages out of the 6 requested (Bengali, Hindi, Malayalam, Tamil and Telugu). We describe the process followed to build the voices from given data and the architecture of our system. In particular, we introduce a penalty in the concatenation cost, inherited from the field of corpus covering, in order to block some concatenations based on their phonological class. Moreover, a fuzzy function is proposed to relax the penalty based on the concatenation quality with respect to the cost distribution. Considering that no language specific processing has been done and that we had never worked with Indian languages, results are very satisfying.