In speech recognition systems, CHMM (Continuous Hidden Markov Model) based speech recognition algorithms have the best accuracy but with the most computational cost. Neither General Purpose Processor (GPP) nor dedicated hardware implementation is a good solution for the algorithm, due to high power consumption for the former and lack of flexibility for the later. To reduce power consumption and enhance flexibility, this paper presents a speech recognition system composed of a coprocessor and a MCU. The coprocessor is a dedicated hardware design for Output Probability Calculation (OPC), which is the most computation-intensive part in CHMM, and MCU is a 32bit RISC (ARM). Tested with a 358-state 3-mixture 27-feature 800-word HMM, MCU operates at 40MHz and coprocessor operates at 10MHz to meet real-time requirement. The power consumption of MCU is 10mW, and coprocessor 1.8mW.