Modification of Sphinx 3 for embedded system implementation

Previous work has determined that the mgau_eval function that calculates the senone score in the Sphinx 3 speech recognition system is the most time consuming, typically requiring up to half of the total execution time. The available version of this function uses floating point operations which are notoriously slow to execute in embedded environments. In this paper, the source code of mgau_eval is analyzed and two new scaled integer versions are developed and implemented. The comparative performances of three versions of mgau_eval when executed on a PC workstation and on an ARM9 processor are presented. The reduction in execution time of the new scaled integer versions are up to 92% on the embedded platform.