Numerous authentic speech samples should be played to learners in Computer-Assisted Language Learning (CALL) when the material is aimed at improving their listening and/or speaking abilities. However, speech files are often compressed in materials that deliver content over the network, such as e-Learning systems, to keep the transfer speed high or to conserve memory space. Therefore, there is a possibility that training with compressed speech will decrease its effects compared to training with uncompressed speech. The effects of speech format on the intelligibility of English phoneme pairs, which are difficult for Japanese to distinguish, were measured through perception experiments. Both native speakers of American English (AE) and native speakers of Japanese (J) identified English word pairs that contrasted /r/-/l/, /b/-/v/, or /s/-/θ/. The results demonstrated that the intelligibility of certain phonemes in certain formats decreased, both for native speakers of AE and native speaker of J. We also found that when native speakers of J were trained to identify English phonemes by using speech with insufficient intelligibility, the training effects could not be generalized to identify high-quality speech. These results suggest that there are some speech formats that are inadequate for use in foreign-language education materials.