A hybrid acoustic-phonetic decoder was developed to recognize intervocalic stop consonants. This recognizer combines the advantages of both explicit acoustic-phonetic knowledge and neural networks. It has been trained and tested on three corpora of continuous speech multispeakers. The statistical tests indicated that the recognition order was: %k (93%) > %t (87%) > %p (84%), and that the three stops were better identified in back and rounded front contexts. To evaluate the efficiency of the acoustic-phonetic knowledge, we compared the hybrid recognizer results to those of a neural recognizer which uses cepstrum coefficients (MFCC). The statistical tests indicated that the former were generally better than the latter.