Abstract: Discusses the combination of phone-level confidence measures using SVM (support vector machines) in speech recognition. Phone-level confidence scores are computed using transitional method while word-level confidence scores are computed from phone-level confidence scores using support vector machines. Experiments on speaker-independent mandarin isolated word recognition showed that rEER (equal error rates,EER) with the proposed method are lower than those with traditional methods (the rEER dropped from 28.14% in baseline to 23.71% in the proposed method), and the complexity of the system increased only slightly.