If you are only using Soundex then you will get all the records that have the same Soundex - it doesn't matter what language you use.
Without showing us any of the "unnecessary" values that are being returned it is difficult to advise any specific improvements.
However, you might want to apply a Levenshtein distance algorithm to the soundex results to further filter them down e.g.
Levenshtein Distance Algorithm – SQLServerCentral[
^]