The goal of this work is the automatic alignment of a map task dialog corpus collected for European Portuguese. The Coral corpus has been orthographically labeled, however off-the-shelf alignment techniques do not work because of the large amount of cross-talk and pronunciation variation. This paper addresses these two issues. The cross-talk problem is dealt with by using a pre-processing stage of channel separation, which proved specially advantageous in the alignment of overlapping speaker turns. The pronunciation variation problem was addressed by including alternative pronunciation rules in the alignment procedure. The alignment is based on WFSTs in the sense that its search space is defined by a distribution-to-word (or distributionto- phone) transducer. Despite many limitations, such as the inadequacy of our current acoustic phone models in terms of voice quality changes (such as laughing), the aligner proved sufficiently robust and demonstrated the feasibility of our alternative pronunciation rules implementation.