Learning Sub-Word Units for Open Vocabulary Speech Recognition

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 712–721, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Learning Sub-Word Units for Open Vocabulary Speech Recognition Carolina Parada1, Mark Dredze1, Abhinav Sethy2, and Ariya Rastrow1 1Human Language Technology Center of Excellence, Johns Hopkins University 3400 N Charles Street, Baltimore, MD, USA carolinap@jhu.edu, mdredze@cs.jhu.edu, ariya@jhu.edu 2IBM T.J. Watson Research Center, Yorktown Heights, NY, USA asethy@us.ibm.com Abstract Large vocabulary speech recognition systems fail to recognize words beyond their vocab- ulary, many of which are information rich terms, like named entities or foreign words. Hybrid word/sub-word systems solve this problem by adding sub-word units to large vo- cabulary word based systems; new words can then be represented by combinations of sub- word units. Previous work heuristically cre- ated the sub-word lexicon from phonetic rep- resentations of text using simple statistics to select common phone sequences. We pro- pose a probabilistic model to learn the sub- word lexicon optimized for a given task. We consider the task of out of vocabulary (OOV) word detection, which relies on output from a hybrid model. A hybrid model with our learned sub-word lexicon reduces error by 6.3% and 7.6% (absolute) at a 5% false alarm rate on an English Broadcast News and MIT Lectures task respectively. 1 Introduction Most automatic speech recognition systems operate with a large but limited vocabulary, finding the most likely words in the vocabulary for the given acoustic signal. While large vocabulary continuous speech recognition (LVCSR) systems produce high quality transcripts, they fail to recognize out of vocabulary (OOV) words. Unfortunately, OOVs are often infor- mation rich nouns, such as named entities and for- eign words, and mis-recognizing them can have a disproportionate impact on transcript coherence. Hybrid word/sub-word recognizers can produce a sequence of sub-word units in place of OOV words. Ideally, the recognizer outputs a complete word for in-vocabulary (IV) utterances, and sub-word units for OOVs. Consider the word “Slobodan”, the given name of the former president of Serbia. As an un- common English word, it is unlikely to be in the vo- cabulary of an English recognizer. While a LVCSR system would output the closest known words (e.x. “slow it dawn”), a hybrid system could output a sequence of multi-phoneme units: s l ow, b ax, d ae n. The latter is more useful for automatically recovering the word’s orthographic form, identify- ing that an OOV was spoken, or improving perfor- mance of a spoken term detection system with OOV queries. In fact, hybrid systems have improved OOV spoken term detection (Mamou et al., 2007; Parada et al., 2009), achieved better phone error rates, espe- cially in OOV regions (Rastrow et al., 2009b), and obtained state-of-the-art performance for OOV de- tection (Parada et al., 2010). Hybrid recognizers vary in a number of ways: sub-word unit type: variable-length phoneme units (Rastrow et al., 2009a; Bazzi and Glass, 2001) or joint letter sound sub-words (Bisani and Ney, 2005); unit creation: data-driven or linguistically motivated (Choueiter, 2009); …

Content truncated to 500 words