This paper investigates the problem of incorporating auxiliary information (e.g. pitch) for speech recognition using dynamic Bayesian networks (DBNs). Previous works usually model acoustic features conditional on the pitch auxiliary variable for both voiced and unvoiced phonetic states, and therefore ignore the fact that pitch (frequency) information is meaningful only for voiced states.. In this paper we propose a switching two auxiliary chain model tailored to voiced/unvoiced states for exploiting pitch information, which is essentially built on the switching parent functionality of Bayesian multinets. Experiments on the OGI Numbers database show that significant performance improvements are achieved from switching auxiliary chain modeling, compared with regular auxiliary chain modeling and the standard HMM.