Since the tangents are flat though could you not simply average it out and simply take the magnitude of the vector describing the translation from key1 to key2 and then divide that by the time between the keyframes?

[ie: so you're assuming a linear transformation over time?]