The paper addresses the issue whether the "probabilities" delivered by a speech recognizer can be directly used as a measure for the confidence of the recognition. As current recognizers have to commit a lot of modelling assumptions and because of estimation problems due to sparse data this certainly is questionable. Nevertheless, this investigation shows, in the framework of recognizing semantic items in the Philips automatic telephone exchange board system PADIS, that there exists a useful correlation between probabilities and confidences. The method proposed works out as a generalization of the more standard method of using likelihood ratios between the first- and second-best recognition path. It offers as distinct advantages a) the integration of all available knowledge sources, and b) the direct and theoretically sound computation of confidence measures on all levels of interest.