Nishimura (2003, 2006) pointed out that in Japanese loanwords, voiced obstruent geminates can optionally devoice when they co-occur with another voiced obstruent (e.g. /doggu/ → [dokku] ‘dog’). This devoicing pattern has been analyzed within a number of theoretical frameworks, and has thereby contributed much to address several theoretical issues. The pattern, moreover, has been studied in several experimental, judgment-based studies. However, there are only a few studies on actual production data. Moreover, all of the previous studies have generally assumed that the devoicing pattern under question is a sociolinguistically monolithic phenomenon. This paper addresses these two issues. By studying the Corpus of Spontaneous Japanese (Kokuritsu-Kokugo-Kenkyuujo, 2008), we first confirm the previous claim that the OCP makes devoicing of geminates more likely in actual production data. Moreover, the results also reveal that many external, sociolinguistic factors affect the applicability of devoicing. The overall results thus contribute to the deeper understanding of the phenomenon by revealing various hitherto unnoticed factors that affect the applicability of devoicing.