In this study, we investigate methods of (a) detecting phonetic boundaries directly from acoustics, and (b) integrating these into HMM-based speech recognition. We test the hypothesis that detecting phone boundaries may be easier using phonological features rather than phonetic or direct acoustic information. We also show how HMMs can be more attuned to the transition of phone boundaries by explicitly modeling transition states. Using a 5-state HMM phone model, we improve the accuracy of phone recognition on the TIMIT task.