Implementation of speech recognition algorithm for a 32-bit CPU-based portable device

Abstract

We have implemented a connected word speech recognition algorithm for a handheld educational device, SpeakingPartner. This system is based on a low-cost 32-bit CPU, and equips a graphic LCD for animation, 8 Mbyte of DRAM and 32 MB of NAND flash memory. Several previously known software optimization algorithms, such as fixed-point conversion, loop fusion, loop unrolling, and circular addressing, as well as efficient log-likelihood computation and partial backtracking methods are applied. The implementation results show that the peak computation requirement for a real-time implementation is reduced to about 8 MIPS, which translates that a 60 MHz 32-bit CPU based system can perform the speech recognition while playing animation in a multi-tasking mode.