We are improving a flexible, large vocabulary, speaker independent, isolated-word recognition system in a telephone environment, originally designed as an integrated system doing all the recognition process in one step. We have transformed it, by adopting the hypothesis-verification paradigm. In this paper, we will describe the architecture and results of the hypothesis subsystem. We will show the system evolution and the modifications adopted to face such a difficult task, achieving significant improvements using automatically clustered phoneme-like units, semi-continuous HMMs, and multiple models per unit. Also, system behavior for vocabulary dependent and independent tasks and vocabularies up to 10000 words will be tested.