Abstract

The high cost of the infrastructure required to conduct state-of-the-art speech recognition research prevents many small research groups from evaluating new ideas on large-scale tasks. To overcome this barrier, we are developing an Internet-based speechto -text (STT) toolkit. In this paper, we present the core component of this system: a decoder that uses a one-pass time-synchronous Viterbi-based search algorithm called trace projection. This decoder can support efficient lattice rescoring using cross-word triphones, lexical trees and n-gram grammars. The decoder performance in terms of CPU and memory usage is on par with commercial systems of its kind. Preliminary evaluations on the SWITCHBOARD (SWB) corpus have yielded a word error rate of 39%. 1. INTRODUCTION A speech-to-text (STT) system conceptually consists of three subsystems --- an acoustic processor which converts the speech signal into a sequence of feature vectors modeled using Hidden Markov Models (HMMs); a linguistic proc...

Citations