I think you are going to have to specify your problem more clearly, what exactly are you trying to achieve with this metric?
I say this, because Levenstein works since it maps pairs of strings to a metric, which can preserve the dimensionality of the string space. What happens if you try and map strings to numbers is that there is a large loss of dimensional information. For example, say I have the string "cat", I'd want "bat", "hat", "rat", "can", "cot" etc. to all be reasonably close to this. With a large number of words, the result is that you end up with dissimilar words being close relatively often e.g. "bat" and "cot" may be close, because they both happen to be similar distances from "cat" on the positive side. This is similar to the problem of what happens when you try and map the plane to a line, it is difficult to meet the restriction that points far away in the plane stay far away on the line. So, the upshot of this is that the 'The more "different" two given strings are, the more different two corresponding values should be' requirement is difficult.
So, my first suggestion is, do you really need something that does this, will a simple hash-code suffice to give you unique values, or perhaps you can use Levenstein after all and ignore the values for individual strings? If none of those suffice, perhaps you can use a multidimensional function value, that is map strings into pairs, triples or another small tuple of numbers. The extra dimensionality granted that way will give you far better results.
An example might be encoding the string as a triple: length, sum of values of letters in string, alternating sum of values of letters e.g. f("cat") = (3, 3 + 1 + 20, 3 - 1 + 20) = (3, 24, 22). This would have some of the properties you desire, but is probably not optimal. Try looking for orthogonal features of the string to do this encoding, or even better, if you have a large test set of strings there are existing libraries for mapping this sort of data into low dimensions while preserving metrics (e.g. the Levenstein metric) and you can train your function on that. I remember the S language had support for this sort of thing.