Javanese is an Indonesian culture which needs to be preserved, but many Javanese students make mistakes in the pronunciation of Javanese letters and find it difficult to analyze errors by human teachers because of the limited time and subjective assessment, so a system is needed to detect incorrect pronunciation of Javanese letters. Mispronunciation detection system has been widely applied in foreign languages, but the system has not been implemented for Javanese carakan letters. This research develops the Javanese letters mispronunciation detection system using Back-Propagation Artificial Neural Networks (BP-ANN). The dataset is obtained from the recorded pronunciation of hanacaraka texts by 24 speakers with 5 repetitions. ALNS method then used to automatically segment the signal into syllables. ANN-PB use statistical value of Mel-Frequency Cepstral Coefficient (MFCC) method with 7 and 14 coefficients. 10-Fold Cross Validation is used to validate and test the system. The Javanese mispronunciation detection using 7MFCC coefficients produces the highest accuracy of 80,07%. While the Javanese mispronunciation detection using 14 MFCC coefficients produces an accuracy of 82.36% at the highest.