In this paper we discuss a data driven approach to select better phone model topologies, in particular to decide on the number of states for linear left-right continuous HMMs. The novel approach is based on a conditional probabilistic viterbi path estimation and operates on forward-backward trained multiple parallel-path HMMs consisting of two different topologies. We compare this conditional probabilistic viterbi path estimation with systematic, statistical and knowledge based designs of different monophone based continuous HMM topologies and evaluate them in a LVSRS system with speech data from the German Verbmobil corpus. We obtained a 5.8% of WER reduction compared to a system with uniform three-state length topologies and a 1.8% WER reduction compared with results of a knowledge based approach with heterogeneous selection of emitting HMM states.