The assistive hearing devices are the only aids to help subjects with hearing loss to use their residual hearing. However, the performance of those devices is closely dependent on auditory training. To develop handheld devices based personal auditory training system with perceptional discrimination analysis and automatic test item generation is very helpful for subjects with hearing loss. Besides, it would ease the burden of speech-language pathologists in developing a personal auditory training. In this study, the mel-frequency cepstrum coefficients and automatic speech recognition are applied to objectively estimate the phonemic confusions. For reducing computational complex, multidimensional scaling is then used to transfer the phonemic confusions into a Euclidean space. Thus, a suitable training material could be automatically generated by simple random process. Finally, the Android based mobile phones are selected as a platform for auditory training. It is convenient for subjects to use the auditory training system. The experimental results show that the average score of mean opinion score is 3.73, which means that the system is very useful.