Abstract

Abstract. Speech recognition is a computationally demanding task, particularly the stage which uses Viterbi decoding for converting pre-processed speech data into words or sub-word units. Any device that can reduce the load on, for example, a PC's processor, is advantageous. Hence we present FPGA implementations of the decoder based alternately on discrete and continuous hidden Markov models (HMMs) representing monophones, and demonstrate that the discrete version can process speech nearly 5,000 times real time, using just 12% of the slices of a Xilinx Virtex XCV1000, but with a lower recognition rate than the continuous implementation, which is 75 times faster than real time, and occupies 45% of the same device.

Abstract. Performing Viterbi decoding for continuous real-time speech recognition is a highly computationally-demanding task, but is one which can take good advantage of a parallel processing architecture. To this end, we describe a system which uses an FPGA for the decoding and a PC for pre-and post-processing, taking advantage of the properties of this kind of programmable logic device, specifically its ability to perform in parallel the large number of additions and comparisons required. We compare the performance of the FPGA decoder to a software equivalent, and discuss issues related to this implementation.

We present a hardware implementation of the log-add algorithm, being a simple method of computing ln(A + B) given ln(A) and ln(B), as used in speech recognition.We show that it can be efficiently implemented in hardware using a small look-up table plus some additional arithmetic logic, with no significant loss of accuracy over direct calculation.

No abstract