By examining acoustic marks of Czech stress, this paper attempts to provide an answer to the question of whether or not perceived accents in the Czech language have an objective existence. A neural network is used to predict the position of accents without lexical information. Three parameters (intonation, duration and intensity) are considered individually, in pairs and altogether. Fundamental frequency seems to be the best predictor of stress, both alone and combined with other parameters. The analysis of the individual prediction errors allows for a closer look at factors which are critical in accent prediction.