Recently, automatic speech recognition method which determines the best structure of HMM by GA was proposed. Availability of the method has been shown. This method was extended to be a new method in which an individual was defined as a collection of the categories. However, better recognition structure was hard to be found because the rate of death-gene increased from one out of thirty in the former method to eleven in the extended method, then search area became narrow. Therefore we revise the method and adopt the death-gene removal algorithm. We show the result of our experiment showing the effectiveness of the revised method.