I haven't tried it myself, but this PLOS ONE paper also shows a possible workflow:

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0066003

If you only care about the human gene expression, then I would treat the mouse reads as contaminants. Bowtie and CLC Bio both allow you to create .fastq files for unaligned reads (so, I would filter out reads alignable to the mouse genome, and then analyze the remaining reads just like a normal human RNA-Seq experiment). If I accidentally align a mouse sample to the human genome (or vice versa), the alignment rate isn't very good - so, I think there should be a fair number of unambiguous reads (which also what the PLOS ONE paper indicates - in fact, those samples had a lot more human reads than mouse reads to begin with).