We present a new acoustic model for speech recognition that explicitly accounts for information omitted from current acoustic models: the definitions of derived acoustic features such as estimated derivatives. We incorporate this information using the method of maximum entropy. We find that, compared to the corresponding HMM, our model cuts an already low error rate about in half for a simple task. We also examine the consequences of ignoring the origin of derived features in CDHMM systems, showing that such an omission in the acoustic model can severely distort the effective language model.