This paper deals with improved confidence assessment for detecting recognition errors and out of dictionary names in a Spanish Recognizer of continuously spelled names over the telephone. We present a hypothesis-verification approach for spelled name recognition. We evaluate the system for several dictionaries, obtaining more than 90.0% recognition rate for a 10,000 name dictionary. For confidence scoring, we consider several features obtained from the different recognition stages. The paper investigates the ability of each feature set to detect recognition errors and names out of the spelling dictionary. We use a neural network to combine all the features in order to obtain the best confidence annotation. Using the data collected from 1,000 phone calls, it is shown that 57.9% incorrectly recognized names and 68.3% out of the spelling dictionary names are detected at a 5% false rejection rate. 1.