Sub-lexical language models for vocabulary speech recognition

摘要

The current practice in speech recognition is to restrict the vocabulary to the most frequent words in the domain. The presence of out-of-vocabulary (OOV) words is, however, inevitable. Since most OOV words are semantically important, recently there has been a surge of research aimed not only at detecting and isolating OOV words, but also at transcribing them. Success, so far, has only been moderate, even on very constrained tasks. We present a new approach to unlimited vocabulary speech recognition based on using grapheme 4o-phoneme correspondences for sublexical modeling of OOV words, and also some very encouraging results we obtained with our approach on a large vocabulary speech recognition task.