Vocabulary Decomposition for Estonian Open Vocabulary Speech Recognition

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 89–95, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Vocabulary Decomposition for Estonian Open Vocabulary Speech Recognition Antti Puurula and Mikko Kurimo Adaptive Informatics Research Centre Helsinki University of Technology P.O.Box 5400, FIN-02015 HUT, Finland {puurula, mikkok}@cis.hut.fi Abstract Speech recognition in many morphologi- cally rich languages suffers from a very high out-of-vocabulary (OOV) ratio. Earlier work has shown that vocabulary decomposition methods can practically solve this problem for a subset of these languages. This pa- per compares various vocabulary decompo- sition approaches to open vocabulary speech recognition, using Estonian speech recogni- tion as a benchmark. Comparisons are per- formed utilizing large models of 60000 lex- ical items and smaller vocabularies of 5000 items. A large vocabulary model based on a manually constructed morphological tag- ger is shown to give the lowest word er- ror rate, while the unsupervised morphol- ogy discovery method Morfessor Baseline gives marginally weaker results. Only the Morfessor-based approach is shown to ade- quately scale to smaller vocabulary sizes. 1 Introduction 1.1 OOV problem Open vocabulary speech recognition refers to au- tomatic speech recognition (ASR) of continuous speech, or “speech-to-text” of spoken language, where the recognizer is expected to recognize any word spoken in that language. This capability is a re- cent development in ASR, and is required or benefi- cial in many of the current applications of ASR tech- nology. Moreover, large vocabulary speech recogni- tion is not possible in most languages of the world without first developing the tools needed for open vocabulary speech recognition. This is due to a fun- damental obstacle in current ASR called the out-of- vocabulary (OOV) problem. The OOV problem refers to the existence of words encountered that a speech recognizer is unable to recognize, as they are not covered in the vocabu- lary. The OOV problem is caused by three inter- twined issues. Firstly, the language model training data and the test data always come from different samplings of the language, and the mismatch be- tween test and training data introduces some OOV words, the amount depending on the difference be- tween the data sets. Secondly, ASR systems always use finite and preferably small sized vocabularies, since the speed of decoding rapidly slows down as the vocabulary size is increased. Vocabulary sizes depend on the application domain, sizes larger than 60000 being very rare. As some of the words en- countered in the training data are left out of the vo- cabulary, there will be OOV words during recogni- tion. The third and final issue is the fundamental one; languages form novel sentences not only by combining words, but also by combining sub-word items called morphs to make up the words them- selves. These morphs in turn correspond to abstract grammatical items called morphemes, and morphs of the same morpheme are called allomorphs of that morpheme. The study of these facets of language is aptly called morphology, and has been largely ne- glected in …

Content truncated to 500 words