Given the availability of complete genome sequences from related organisms, sequence conservation can provide important clues for predicting gene structure. In particular, one should be able to leverage information about known genes in one species to help determine the structures of related genes in another. Such an approach is appealing in that high-quality gene prediction can be achieved for newly sequenced species, such as mouse and puffer fish, using the extensive knowledge that has been accumulated about human genes. This article reports a novel approach to predicting the exon-intron structures of mouse genes by incorporating constraints from orthologous human genes using techniques that have previously been exploited in speech and natural language processing applications. The approach uses a context-free grammar to parse a training corpus of annotated human genes. A statistical training procedure produces a weighted recursive transition network (RTN) intended to capture the general features of a mammalian gene. This RTN is expanded into a finite state transducer (FST) and composed with an FST capturing the specific features of the human orthologue. This model includes a trigram language model on the amino acid sequence as well as exon length constraints. A final stage uses the free software package ClustalW to align the top n candidates in the search space. For a set of 98 orthologous human-mouse pairs, we achieved 96% sensitivity and 97% specificity at the exon level on the mouse genes, given only knowledge gleaned from the annotated human genome.