We asked whether the identification of emotional facial expressions (FEs) involves the simultaneous perception of the facial configuration or the detection of emotion-specific diagnostic cues. We recorded at high speed (500 frames s−1) the unfolding of the FE in five actors, each expressing six emotions (anger, surprise, happiness, disgust, fear, sadness). Recordings were coded every 10 frames (20 ms of real time) with the Facial Action Coding System (FACS, Ekman et al 2002, Salt Lake City, UT: Research Nexus eBook) to identify the facial actions contributing to each expression, and their intensity changes over time. Recordings were shown in slow motion (1/20 of recording speed) to one hundred observers in a forced-choice identification task. Participants were asked to identify the emotion during the presentation as soon as they felt confident to do so. Responses were recorded along with the associated response times (RTs). The RT probability density functions for both correct and incorrect responses were correlated with the facial activity during the presentation. There were systematic correlations between facial activities, response probabilities, and RT peaks, and significant differences in RT distributions for correct and incorrect answers. The results show that a reliable response is possible long before the full FE configuration is reached. This suggests that identification is reached by integrating in time individual diagnostic facial actions, and does not require perceiving the full apex configuration.