Development of a WFST based Speech Recognition System for a Resource Deficient Language Using Machine Translation
Arnar Thor Jensson, Tasuku Oonishi, Koji Iwano, Sadaoki Furui · Hokkaido University Collection of Scholarly and Academic Papers (Hokkaido University) · 2009
Text corpus size is an important issue when building a language model (LM) in particular where insufficient training and evaluation data are available. In this paper we continue our work on creating a speech recognition system with a LM that is trained on a small amount of text in the target language. In order to get better performance we use a large amount of foreign text and a dictionary mapping between the languages. A dictionary is used since we are assuming that the target language is resource deficient and therefore statistical machine translation (MT) is not available. In this paper we take a step forward from our previous published method by using a coupling of the speech recognition part and the translation part rather than pre-translating the foreign text. The coupling is achieved with a weighted finite state transducer (WFST) network which as well makes it possible to easily switch between the output language, i.e. that the output text is in the format of the resource deficient language or in the resource rich language. Our method outperforms the resource-deficient Icelandic speech recognition baseline, 82.6% keyword accuracy (KA), when the system is trained on 1500 Icelandic sentences, both for the English output (2.6% absolute KA improvement) and for the Icelandic output (1.6% absolute KA improvement) where the English text corpus consists of 63003 sentences.