Effects of phonetic transcription on a sub-word based name recognition system

The effects of altering the phonetic transcription on a sub-word based speaker independent name recognition system are reported. A connected word recognition algorithm is used with sub-word models constrained by a grammar and a lexicon. Two approaches are compared. In the first, all names are hand transcribed by a human expert. Both multiple and single phonetic transcription are provided for each name. In the second, an automatic pronunciation generator (APG) is used to translate English words or names to their phonetic transcriptions. Using multiple vs. single transcriptions of a human expert leads to a decrease of 16%-28% in recognition error rates. Multiple transcriptions provided by the APG increases the recognition error rates by an average factor of 2.75 compared to those obtained by the human expert.