This paper describes our methods used in the NAIST-NICT submission to the International Workshop on Spoken Language Translation (IWSLT) 2012 evaluation campaign. In particular, we propose two extensions to minimum bayesrisk decoding which reduces a expected loss.