Integrating connectionist, statistical and symbolic approaches for continuous spoken Korean processing

This paper presents a multi-strategic and hybrid approach for large-scale integrated speech and natural language processing, employing connectionist, statistical and symbolic techniques. The developed spoken Korean processing engine (SKOPE) integrates connectionist TDNN-based phoneme recognition technique with statistical Viterbi-based lexical decoding and symbolic morphological/phonological analysis techniques. The modular large-scale TDNNs are organized to recognize all 41 Korean phonemes using 10 component networks combined through 3 glue networks. In performance phase, continuously shifted TDNN outputs are integrated with HMM-based Viterbi decoding using a tree-structured lexicon. The Viterbi beam search is integrated with Korean morphotactics and phonological modeling, and produces a morpheme-graph for high-level parsing module. Currently. SKOPE shows average 76.2% phoneme spotting performance for all 41 Korean phonemes (including silence) from continuous speech signals and exhibits average 92.6% morpheme spotting performance from erroneous TDNN outputs after morphological analysis. Other extensive experiments verify that the multi-strategic approaches are promising for complex integrated speech and natural language processing, and the approaches can be extended to other morphologically-complex agglutinative languages such as Japanese