In recent years, a great deal of research on second language (L2) acquisition has been concerned with non-target production of L2 learners, addressing issues such as native language (L1) transfer into L2 and the nature and source of developmental errors. Previous studies have mostly focused on the analysis of discrete L2 segments (Flege 1987, 1999; Major & Kim 1996), rather than on L2 phonological patterns. This study, however, examines the production of sequences of sounds in Russian L1 and L2 from both the phonetic and phonological perspectives.This dissertation investigates native and non-native production of real and nonsense words containing obstruent clusters in which a phonological phenomenon known as regressive voicing assimilation is required. In Russian, forms like lodka `boat' are rendered orthographically with a voiced obstruent which is pronounced as a voiceless one when followed by a voiceless obstruent. The results of the experiments reveal several production patterns in L1 and L2 speech as well as gradiency in devoicing which are further analyzed within the stochastic Optimality Theory framework. Categorical production is accounted for by the re-ranking of L1 and L2 constraints; whereas, gradiency in production is viewed as a result of the re-ranking of constraints within phonetically detailed constraint families.