This paper describes a system for accurate continuous speech recognition called "ATREUS/SSS-LR". A phoneme-context-dependent LR parser drives allophonic HMMs represented by a shared-state network automatically generated by the Successive State Splitting (SSS) algorithm. In this system, the SSS principle has also been applied to duration clustering: optimal clusters of phoneme-context-dependent durations are automatically generated independently of the HMnet-based allophonic classes.

ATREUS/SSS-LR achieved a phrase recognition rate of 93.2%, the best recognition result achieved in the 1000-word recognition experiments conducted at ATR. This recognition rate was obtained with a smaller beam width than used with discrete HMM(fuzzy vector quantization) and continuous mixture density HMM. This shows that the SSS-LR can realize both fast parsing and high accuracy.