Third International Conference on Spoken Language Processing (ICSLP 94)
In this paper, we propose a spoken Korean morphological analysis model extensible to large vocabulary continuous speech recognition. This model consists of a diphone recognizer, a diphone2phoneme filter and a CYK-morphological analyzer. Two-level hierarchical TDNNs (time-delayed neural networks) recognize Korean diphones which are transformed into a phoneme lattice (a set of phoneme candidates hypothesized by a speech recognition module) by a diphone2phoneme filter. The morphological analyzer parses the phoneme lattice with the phonological changes handling and produces the morphology-segmented Korean words (called Eojeols). Using the TDNN diphone speech recognizer, we obtained 95.2% of 17 Korean vowel recognition and 93.7% of 72 diphone recognition. The speaker-dependent and continuous Eojeol recognition experiments using the current model show that the morphological analysis for spoken Korean can be achieved for medium sized vocabularies with 90.6% of success rate.
Bibliographic reference. Kim, Kyunghee / Lee, Geunbae / Lee, Jong-Hyeok / Jeong, Hong (1994): "Integrating TDNN-based diphone recognition with table-driven morphology parsing for understanding of spoken Korean", In ICSLP-1994, 5-8.