A post-processing system to yield reduced word error rates: Recognizer Output Voting Error Reduction (ROVER)

TLDR

NIST developed a composite ASR system that combines outputs from multiple recognizers, often achieving lower error rates than any single system, and adding more knowledge sources generally reduces errors. This paper presents a post‑recognition process that treats multiple ASR outputs as independent knowledge sources to produce a lower‑error composite output. The ROVER system merges ASR outputs into a minimal‑cost word transition network using dynamic‑programming alignments, then applies a voting/automatic rescoring step to select the sequence with the lowest score.

Abstract

Describes a system developed at NIST to produce a composite automatic speech recognition (ASR) system output when the outputs of multiple ASR systems are available, and for which, in many cases, the composite ASR output has a lower error rate than any of the individual systems. The system implements a "voting" or rescoring process to reconcile differences in ASR system outputs. We refer to this system as the NIST Recognizer Output Voting Error Reduction (ROVER) system. As additional knowledge sources are added to an ASR system (e.g. acoustic and language models), error rates are typically decreased. This paper describes a post-recognition process which models the output generated by multiple ASR systems as independent knowledge sources that can be combined and used to generate an output with reduced error rate. To accomplish this, the outputs of multiple of ASR systems are combined into a single, minimal-cost word transition network (WTN) via iterative applications of dynamic programming (DP) alignments. The resulting network is searched by an automatic rescoring or "voting" process that selects the output sequence with the lowest score.