Hidden Markov model (HMM) and recurrent neural net recognizers were used to classify phonetic segments from the TIMIT database. The neural net was derived from the HMM by removing the probability constraints on the parameters and using an error-correcting, minimum-mean-squared-error (MMSE) training criterion. The neural net was 0. 3% to nearly 5% better than the HMM (corresponding to 2% to 15% reductions in error rate), depending on the phonetic class. The improvement was due to both the error-correcting training and the removal of the parameter constraints. Some comparisons were also done with a maximum mutual information (MMI) training criterion; MMI yielded only slight improvements over ML, and was more difficult to train to than was MMSE.