This paper presents a new scoring method for text-dependent speaker recognition. The scoring method is used to combine the scores measured by the phoneme-based neural tree network (NTN) classifiers. In contrast to the conventional method that combines the output scores of speech frames by averaging them over the utterance, the new method uses phoneme-dependent weighted combination. Since the phonemes are different in their effectiveness for speaker discrimination, the phonetic weights are chosen according to their abilities of speaker discrimination. The proposed method is evaluated by experiments on the YOHO database. Performance improvements are obtained over conventional techniques.