The studies reported in this paper have dealt with the responses of human observers to speech stimuli transmitted in a background of white Gaussian noise. In all cases the listeners attempted to identify the transmitted items and then made a second response in an attempt to convey additional information. It was found that when the listeners were allowed a second‐choice identification response, very little information was contained in these responses which was not already contained in the listeners' first identification response. When the second response was a confidence rating, significant amount of information was added to that which was carried by the identification response.
The rating which followed each identification response was assigned by the observers in an attempt to estimate the probability that their identification response was, in fact, correct. For message sets of four items and for sets of sixteen items, it was found that the observers were quite capable of making such estimates over a wide range of speech‐to‐noise ratios. Their estimates did appear to be affected to some extent by the size of the message set and by the speech‐to noise ratio, but this interaction was slight.
The observers' rating responses were used to generate ROC curves. These curves were adequately fit by straight lines when the data were plotted on normal‐normal probability paper. Regardless of the size of the message set, all curves, for all speech‐to‐noise ratios, were fit by a single slope. However, the point at which these curves intersected the abscissa was a function of both variables.
Data from one set of observers in the rating experiments were used in an attempt to predict the performance of a different group of observers whose task was to monitor subsets of messages. While predictions were fairly good, discrepancies were noted. An internal check in the monitoring experiment strongly suggests that these discrepancies arose because of differences between the two groups of observers.