A speech recognizer performs speech recognition on a spoken name supplied by a user, producing a list of possible matches and corresponding confidence scores, by comparing a packetized voice stream of a spoken name to a plurality of stored phonemes that represent users' text names. If the top scoring match for a spoken name does not correctly identify the spoken name or if the spoken name's confidence score is below a first threshold, the user name is flagged to the system administrator as having a potential speech recognition problem. The results of the speech recognition are used to suggest names whose spelling may need to be adjusted to resolve the speech recognition problem. During production, a low threshold for rejecting speech recognition results can be adjusted downwards for names that produced low scores during testing. Heuristics are presented for re-testing only a subset of names when the set of names is changed.