In this paper, we compare a large-vocabulary speech decoder, based on a phonetic prefix tree, with a decoder based on finite-state transducers. On the example of a large-vocabulary, isolated word recognition task without language model, we investigate the error-rates based on different beamwidths for fully optimized, tree-like and factored finite-state transducer networks. The results show that the decoder, based on a fully optimized finite-state network, achieves the same error-rates as the prefix tree decoder, using roughly 50% of active states. In addition, we modify the traditional phonetic prefix tree to a new tree, compatible with standard LVCSR decoders but optimized across all hidden Markov models to remove redundancies, which achieves better error-rates at small beamwidths.