Charith Mendis1, Geoffrey Zweig2, Jasha Droppo2, Madanlal Musuvathi2, Saeed Maleki2, Todd Mytkowicz2
13:30 - 15:30 | Tue 22 Mar | Poster Area I | SP-P2.2
The performance intensive part of a large-vocabulary continuous speech recognition system is the Viterbi computation that determines the sequence of words that are most likely to generate the acoustic-state scores extracted from an input utterance. This paper presents an efficient parallel algorithm for Viterbi. The key idea is to partition the per-frame computation among threads to minimize inter-thread communication despite traversing a large irregular acoustic and language model graphs. Together with a per-thread beam search, load balancing language-model lookups, and memory optimizations, we achieve a 6.67× speedup over an highly-optimized production-quality WFST-based speech decoder. On a 200,000 word vocabulary and a 59 million ngram model, our decoder runs at 0.27× real time while achieving a word-error rate of 14.81% on 6214 labelled utterances from Voice Search data.