Word alignment issues in ASR scoring
George R. Doddington · 2004
ASR evaluation requires that system output words be aligned with reference words in order to be scored. Alignment typically is done so as to minimize the resulting word error rate (WER), and this alignment process does not consider whether corresponding reference and system output words actually correspond in time. The resulting alignment thus masks some errors and hinders the understanding of ASR failure mechanisms. This paper presents some alignment timing statistics and then shows some characteristics of ASR failures that are missed by typical ASR scoring algorithms.