Integration of multiple speech recognition modules based on RT-middleware

This paper proposes an integrated system of multiple speech recognition modules in order to improve recognition rate. This algorithm integrates recognition result from multiple, homogenous speech recognition modules, each of which is located in different position. The recognition rate of the module at each position is measured in advance, based on which the recognition results of multiple modules in different positions are integrated. In order to make the system flexible and robust, it is designed based on RT-middleware, in which speech recognition module implemented as a RT-component is employed. The system is implemented so that it can work even if there exist broken modules or those that do not respond. Experimental results show that this algorithm has improved recognition rate by integrating result of multiple modules, compared with single module. It is also shown that the system is robust against breakdown or no response of modules.