This paper compares and contrasts the keyword spotting performance of conventional maximum likelihood trained HMMs vs. that of minimum error trained HMMs. The unique aspect of this work is the use of a new minimum error classification algorithm [1] for training the continuous mixture density HMM components of an HMM-based keyword spotting system. The actual spotting algorithm used was the HMM garbage model approach proposed previously in [2] [3]. Speaker-independent keyword spotting experiments were performed using the ATR Japanese continuous speech database. The reported results show the clear superiority of the minimum error trained HMMs in the chosen keyword spotting application.