Theories of word production rely heavily on semantic context effects. For example, when pictures are named repeatedly and sorted by semantic category, it takes longer to name them than when they are named in an intermixed manner. Wöhner et al. (2021) found that semantic interference in blocked-cyclic naming was approximately three times greater for sounds (e.g., barking) than for pictures (e.g., a dog). This study explored possible reasons for this differential interference pattern. When the stimuli were presented briefly, the differential interference was replicated, ruling out the possibility that it results from a large overlap between stimulus processing and phonological response preparation. When the sounds and pictures were replaced with spoken and written words that the participants translated into their second language, the size of the interference was similar, showing that the differential pattern does not result from different stimulus modalities (auditory vs. visual). When participants typed their responses, the differential interference was replicated again, showing that it does not depend on a certain response modality (spoken vs. written). Additionally, pupil dilation data suggest that sounds, compared with pictures, may activate additional representations that contribute to the interference. Overall, the data imply that semantic interference in word production is shaped by early, stimulus-specific, perceptual and semantic “lead-in” processes that precede lexical selection. More attention should be given to these early processes in word production research.