The nature of the self-monitoring system and the nature of the code being monitored in language production Niels O. Schiller Self-monitoring of action is important for smooth performance in many areas of human behavior. For instance, when we reach out to grasp for an object, we are able to monitor our arm movement and quickly modify the trajectory in case an obstacle is suddenly encountered. The same is true for producing speech. Speech is largely planned and the selection of meaning, syntax, and word forms is a complex process in which errors might occur. Self-monitoring the planning process might prevent (some of these) errors to surface in overt speech. This is important because speech errors might hamper the fluency of a conversation and they can sometimes be embarrassing. Models of speech production recognized the importance of a monitoring system that helps speakers to convey their messages in an optimal way. In the first part of my paper, I will discuss the nature of the self-monitoring system used in language production. Generally, one can distinguish between production-based and perception-based self-monitoring systems. The most widely accepted theory is Levelt's (1983, 1989) perceptual loop theory. This theory states that monitoring proceeds through one central, capacity-limited perceptual system. I will present ERP data supporting the hypothesis that the monitoring system is capacity-limited. Participants displayed a lower amplitude of the error negativity (ERN) in a phoneme monitoring task when the response deadline was shorter compared to when it was longer. A possible account for this effect is that the monitoring system includes a comparator component that compares the correct response with the actual response. In case of a mismatch, an error signal is generated. However, when there is less time for comparison, the magnitude of a potential error signal decreases. In the second part of my paper, I will discuss the nature of the code that is being monitored in verbal self-monitoring during speech production, i.e. is it phonetic, phonological, or even more abstract in nature? Wheeldon and Levelt (1995) suggested that the code that is self-monitored in speech production is phonological in nature. I will present more empirical data in favor of this hypothesis. Furthermore, I will present new data that bear on so far untested aspects of their hypothesis. For instance, I will show that lexical stress plays an important role in phoneme monitoring. In contrast, morphological boundaries do not play an independent role. This latter finding is predicted if morphological encoding precedes phonological encoding, as in the model of Levelt, Roelofs, and Meyer (1999). This is another piece of evidence in favor of a phonological representation of the utterance being used for internal monitoring. Furthermore, I will present acoustic-phonetic measurements showing that the monitoring latencies are independent from the physical intervals between the target phonemes in the sound signal. I will also present a tentative model to account for the observed monitoring latencies. This model proposes that features like lexical stress or syllable boundaries which need to be encoding at the phonological word level take time and may delay the encoding - and hence monitoring - of later segments. However, following Wheeldon and Levelt (1995), the model assumes that segments continue to become available at the retrieval level even if they cannot yet be inserted at the encoding level. That way, the model is able to account for the general decrease in monitoring latencies for segments occurring later in a word. The model is able to account for a wide range of empirical data from phoneme monitoring experiments.