Directory name retrieval using HMM modeling and robust lexical access

A system to retrieve names in a directory from the pronunciation and the spelling is presented, for telephone quality speech in French. The recognizer is based on hidden Markov modeling of speech units using word models for letters and phone models for name pronunciation. The directory is built automatically from a list of names using a grapheme to phoneme converter for the names and rules for spellings, each entry in the directory consisting in several pronunciation and spelling variants. From the acoustic recognition, the corresponding entry in the directory is found using dynamic alignment of symbol sequences, with insertion, deletion and substitution costs fixed or determined from the training data to take into account acoustic confusability. Experimental results show that a mixed recognition strategy combining both spelling and pronunciation improves the spelling based strategy. The effectiveness of integrating learnt phonetic confusability information during the lexical search is also demonstrated.