This paper reports the results of experiments in speaker-independent, vocabulary-independent speech recognition using Phoneme Decision Tree (PDT) methods. Phoneme-level hidden Markov models (HMMs) are trained on a corpus comprising a balanced set of general spoken English sentences and evaluated on a test corpus of continuously spoken airborne reconaissance reports. Vocabulary-dependent and -independent performance is compared. Experimental results are reported for "PDT pruning", which explore the trade-off between number of models and amount of training material per model, conventional uniform allocation of mixture components to HMM states, and non-uniform allocation of mixture components to states.