We used a tree-trellis type speech recognition system, which was developed in our laboratory. The Hidden Markov Kits [5] was used to learn phone HMMs. Speaker-independent HMM models were prepared using the ATR A-set word database (containing 5240 words for a set) and speech data of 10-males sets. Phone models were also adapted using the ATR B-set (503 sentences uttered by the same speaker). The word bigram model was developed with the Mainichi newspaper articles published during 1993 (about 100,0000 sentences) and 70 speech input sentences.