Statistical learning plays an important role in acquiring the structure of cultural communication signals such as speech and music, which are both perceived and reproduced. However, statistical learning is typically investigated through passive exposure to structured signals, followed by offline explicit recognition tasks assessing the degree of learning. Such experimental approaches fail to capture statistical learning as it takes place and require post-hoc conscious reflection on what is thought to be an implicit process of knowledge acquisition. To better understand the process of statistical learning in active contexts while addressing these shortcomings, we introduce a novel, processing-based measure of statistical learning based on the position of errors in sequence reproduction. Across five experiments, we employed this new technique to assess statistical learning using artificial pure-tone or environmental-sound languages with controlled statistical properties in passive exposure, active reproduction, and explicit recognition tasks. The new error position measure provided a robust, online indicator of statistical learning during reproduction; with little carry over from prior statistical learning via passive exposure and no correlation with recognition-based estimates of statistical learning. Error position effects extended consistently across auditory domains, including sequences of pure tones and environmental sounds. Whereas the average length of reproduced sequences showed significant variability across experiments, and little evidence of being improved by statistical learning, the error position effect was highly consistent for all participant groups, including musicians and non-musicians. We discuss the implications of these results for understanding psychological mechanisms underlying statistical learning, and compare the evidence provided by different experimental measures.