We analyzed a speaker-independent, continuous Spanish speech recognition system for desktop command & control of a Macintosh personal computer. A word error rate of 0.6% was obtained for the System 7.1 Finder navigation task with "clean" speech recorded under semi-controlled conditions in Cupertino, California, and 2.8% for "real world" speech recorded in an open office environment in Mexico City. In obtaining these results, we demonstrated the utility of cepstral normalization as a means of pooling data across speech data collection sites. Finally, data was obtained showing that for the measured tasks, acoustic, phonetic and linguistic data could be leveraged across Spanish dialects.