Design a co-processor for Output Probability Calculation in speech recognition

In the CHMM (Continuous Hidden Markov Model) based speech recognition algorithm, Output Probability Calculation (OPC) is the most computation-intensive part. To reduce power consumption and design cost, this paper presents a custom-designed co-processor to implement OPC. The standard SRAM interface of the co-processor allows it to be controlled by various micro-controllers. The co-processor has been implemented in standard-cell based approach and manufactured in 0.18 mum UMC technology. Tested with a 358-state 3-mixture 27-feature 800-word HMM, the co-processor operates at 10 MHz to meet real-time requirement. The power consumption of this co-processor is 1.6 mW, and the die size is 1.18 mm².