OOV Handling Using Partial Lemma-Based Language Model in LF-MMI Based ASR for Bahasa Indonesia

Abstract

One of the common problems in ASR is the out-of-vocabulary word in an utterance that can degrade the performance of the system. Bahasa Indonesia, as an agglutinative language, uses affixation to generate words from a set of affixes and root words. We propose the use of a partial lemma-based language model (LM) and lexicon that can handle words created from affixation. The partial lemma-based LM and lexicon are created from the original ones using morphology analyzer output as a reference. The experiment shows that using the LM in ASR with LF-MMI cost function gives a better WER when the heuristic to insert inter-word short pause is modified to also consider the affixes.