LINGUISTIC DISSECTION OF SWITCHBOARD-CORPUS AUTOMATIC SPEECH RECOGNITION SYSTEMS

A diagnostic evaluation of eight Switchboard-corpus recognition systems was conducted in order to ascertain whether word-error patterns are attributable to a specific set of linguistic factors. Each recognition system’s output was converted to a common format and scored relative to a reference transcript derived from phonetically hand-labeled data. This reference material was analyzed with respect to ca. forty acoustic, linguistic and speaker characteristics, which in turn, were correlated with recognition-error patterns via decision-trees and other forms of statistical analysis. The most consistent factors associated with superior recognition performance pertain to accurate classification of phonetic segments and articulatoryacoustic features. Other factors correlated with word recognition are syllable structure, prosodic stress and speaking rate (in terms of syllables per second). 1