This study tested the validity of applying a signal detection model to Japanese listeners' /r/-/l/ labeling behavior in one-interval identification task, and the hypothesis that Japanese listeners estimate each token's goodness of fit to Japanese /r/ in labeling it as /r/ or /l/. The identification data obtained by manipulating subjects' response criteria fit the signal detection model. The performance in the identification task was highly predictable based on the goodness of fit judgment, indicating that category goodness information contained in /r/ and /l/ is utilized in deriving underlying sensory decision variables. The data, however, also suggested that these variables are not a direct consequence of the goodness of fit judgment. Various implications of the signal detection model in the investigation of Japanese perception of /r/ and /l/ are discussed.