Devoicing of high vowels (HVD) in Tokyo Japanese applies in two environments—between voiceless consonants, and between a voiceless consonant and a “pause”—and applies variably as a function of a number of factors. The role and definition of “pause” in this process, in terms of a physical pause or prosodic position (word or phrase boundary), remains unclear, as does what is expected when these environments overlap, and why HVD appears to be categorical in some environments and variable in others. This paper addresses three outstanding issues about HVD—the role of “boundary phenomena” (prosodic position and physical pauses), the relationship between the two environments, and the sources of variability in HVD—by examining vowel devoicing in a large corpus of spontaneous Japanese. We use mixed-effects logistic regression to model how boundary phenomena affect the likelihood of devoicing and modulate the effects of other variables, controlling for other major factors, including a measure of gestural overlap. The results suggest that all boundary phenomena jointly affect devoicing rate, and that prosodic phrase boundaries play a key role: variability in HVD looks qualitatively different for phrase-internal and phrase-final vowels, which are affected differently by word frequency, speech rate, and pause duration. We argue the results support an account of HVD as the result of two overlapping vowel devoicing processes, each widely-attested cross-linguistically: devoicing between voiceless consonants, and devoicing before prosodic phrase boundaries. Variability in the application of these two processes can then be partially explained in terms of aspects of phonetic implementation and processing: gestural overlap (Beckman 1996), which often plays a role in reduction processes, and the locality of production planning (Wagner 2012), a recent explanation for variability in the application of external sandhi processes.