Analysis of a large in-car speech corpus