This paper investigates speaker-independent spelling recognition over the telephone using a known dictionary of possible spelled names which provides very useful information. Several ways of using this knowledge are presented and compared: introduction of syntactical constraints at the decoding level, application of a syntactical post-processing of the N best solutions, and finally utilization of a retrieval procedure which looks for the most probable spelled name, knowing the recognized sequence of letters. A large part of the paper is devoted to the combination of the N-best solutions algorithm with the retrieval procedure. The results reported in this article are obtained using a database containing 3,000 utterances of spelled names of French cities recorded from 180 speakers over the telephone network. Application of the retrieval procedure to the first few solutions delivered by the N-best algorithm leads to a 40 % reduction in the spelling error rate for a dictionary of 120 city names, and a 25% reduction for a dictionary of 30,000 town and city names.
Keywords: Markov modeling, N-best solutions, Spelling recognition