Contextual Information Improves OOV Detection in Speech

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 216–224, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Contextual Information Improves OOV Detection in Speech Carolina Parada, Mark Dredze HLTCOE Johns Hopkins University 3400 North Charles Street, Baltimore MD 21210, USA carolinap@jhu.edu mdredze@cs.jhu.edu Denis Filimonov HLTCOE University of Maryland, College Park, MD 20742 USA den@cs.umd.edu Frederick Jelinek HLTCOE Johns Hopkins University 3400 North Charles Street, Baltimore MD 21210, USA jelinek@jhu.edu Abstract Out-of-vocabulary (OOV) words represent an important source of error in large vocabulary continuous speech recognition (LVCSR) sys- tems. These words cause recognition failures, which propagate through pipeline systems im- pacting the performance of downstream ap- plications. The detection of OOV regions in the output of a LVCSR system is typically ad- dressed as a binary classification task, where each region is independently classified using local information. In this paper, we show that jointly predicting OOV regions, and includ- ing contextual information from each region, leads to substantial improvement in OOV de- tection. Compared to the state-of-the-art, we reduce the missed OOV rate from 42.6% to 28.4% at 10% false alarm rate. 1 Introduction Even with a vocabulary of one hundred thou- sand words, a large vocabulary continuous speech recognition (LVCSR) system encounters out-of- vocabulary (OOV) words, especially in new do- mains or genres. New words often include named entities, foreign words, rare and invented words. Since these words were not seen during training, the LVCSR system has no way to recognize them. OOV words are an important source of error in LVCSR systems for three reasons. First, OOVs can never be recognized by the LVCSR system, even if repeated. Second, OOV words contribute to recog- nition errors in surrounding words, which propagate into to later processing stages (translation, under- standing, document retrieval, etc.). Third, OOVs are often information-rich nouns – mis-recognized OOVs can have a greater impact on the understand- ing of the transcript than other words. One solution is to simply increase the LVCSR system’s vocabulary, but there are always new words. Additionally, increasing the vocabulary size without limit can sometimes produce higher word error rates (WER), leading to a tradeoff between recognition accuracy of frequent and rare words. A more effective solution is to detect the presence of OOVs directly. Once identified, OOVs can be flagged for annotation and addition to the system’s vocabulary, or OOV segments can be transcribed with a phone recognizer, creating an open vocabu- lary LVCSR system. Identified OOVs prevent error propagation in the application pipeline. In the literature, there are two basic approaches to OOV detection: 1) filler models, which explicitly represent OOVs using a filler, sub-word, or generic word model (Bazzi, 2002; Schaaf, 2001; Bisani and Ney, 2005; Klakow et al., 1999; Wang, 2009); and 2) confidence estimation models, which use differ- ent confidence scores to find unreliable regions and label them as OOV (Lin et al., 2007; Burget et al., 2008; Sun et al., 2001; Wessel et al., 2001). Recently, Rastrow …

Content truncated to 500 words