Speaker-independent spelling recognition over the telephone

The authors investigate speaker-independent spelling recognition over the telephone using Markov modeling at two levels: one for the recognition of connected letter sequences and one for the retrieval of the word from a known list. A connected-word speech recognizer must be used to deal with natural spellings, and the retrieval procedure has to take into account the insertion and deletion errors as well as the substitution errors. The speech database, recorded from about 180 speakers, contained 6000 sequences (average length of seven letters) corresponding to the spelling of city names, proper names, and random sequences. On the city names test set, before retrieval the letter error rate was 15.9%. Several retrieval procedures are presented and compared. A Markov modeling approach leads to the best performance with a retrieval error rate of 4.5% for a list of 1000 possible names and 12.4% for a list of 30000 town and city names.<>