A decision rule which assigns an unknown signal to the class for which its average distance to theKnearest samples is minimum is derived. It is shown that this decision rule is equivalent to the classical maximum posterior probability rule based on nonparametric estimates of the local class conditional density functions. We have used the rule successfully in an isolated word recognition experiment and have found that it performs better than other nearest neighbor classification rules. Part of the improved performance is due to the selection of the appropriate value ofK. We show how this choice can be tailored to the training set and why the rule is particularly appropriate for our speech recognition work.