Out-of-vocabulary word modeling using multiple lexical fillers

Abstract

In large vocabulary speech recognition, out-of-vocabulary words are an important cause of errors. We describe a lexical filler model that can be used in a single pass recognition system to detect out-of-vocabulary words and reduce the error rate. When rescoring word graphs with better acoustic models, word fillers cause a combinatorial explosion. We introduce a new technique, using several thousand lexical fillers, which produces word graphs that can be rescored efficiently. On a large French vocabulary continuous speech recognition task, lexical fillers achieved an OOV detection rate of 44% and allowed a 23% reduction in errors due to OOV words.