An important aspect of a Computer-Assisted Language Learning (CALL) system for pronunciation acquisition is the automatic detection of mispronunciations. This problem can be formulated as a phone verification task. For each phone to be verified, the system generates a verification score and a decision threshold is applied to accept or reject the pronunciation of that phone. Most verification systems use the HMM phone acoustic models to compute the log posterior probabilities (LPPs) as the verification score. A discriminative back-end using the Support Vector Machine (SVM) can also be applied to the vector of LPPs to further improve the verification performance. This paper investigates the use of a NN/HMM hybrid phone recognizer to obtain the LPP scores. The NN/HMM hybrid framework has been shown to yield superior phone recognition performance over the conventional GMM/HMM based systems. In addition, this paper also examines the use of frame-level phone or state posterior features directly with SVM. Experimental results reported on the TIMIT database show that state-level average posterior features with SVM yielded 9.5% relative Equal Error Rate (EER) improvement over the NN/HMM system.