FPGA Implementation of a Pipelined Gaussian Calculation for HMM-Based Large Vocabulary Speech Recognition

A scalable large vocabulary, speaker independent speech recognitionsystem is being developed using Hidden MarkovModels (HMMs) for acoustic modeling and a Weighted FiniteState Transducer (WFST) to compile sentence, word,and phoneme models. The system comprises a softwarebackend search and an FPGA-based Gaussian calculationwhich are covered here. In this paper, we present an efficientpipelined design implemented both as an embedded peripheraland as a scalable, parallel hardware accelerator. Both architectureshave been implemented on an Alpha Data XRC-5T1, reconfigurable computer housing a Virtex 5 SX95TFPGA. The core has been tested and is capable of calculatinga full set of Gaussian results from 3825 acoustic modelsin 9.03 ms which coupled with a backend search of 5000words has provided an accuracy of over 80%. Parallel implementationshave been designed with up to 32 cores andhave been successfully implemented with a clock frequencyof 133 MHz