In general, high recognition accuracy can be obtained by using CFG, PCFG, n-grams etc. and thus constraining the vocabulary and grammatical constructs. However, a problem exists in recognizing words which have not been previously recorded in these grammars. In thisresearch, we were researching the possibility of using only weak phonotactic constraints apparent in Japanese, e.g. the sequence p/t/a is not "valid" Japanese. To maintain high recognition accuracy in spite of these very weak constraints, 3278 allophone models which were trained by the Successive State Splitting (SSS) algorithm were than adapted to phrase utterance type using the VFS algorithm. Phrase recognition accuracy for 279 phrases (SB3) was 51.3 %, phoneme accuracy was 92.5 %, a surprisingly good result in view of the fact that only a standard One-Pass DP was used.