Analysis of distinctive feature matching with random error generation in a lexical access system

Abstract

A matcher for a distinctive feature-based lexical access system is tested using degraded feature inputs. The input speech comprises 16 conversation files from a map task in American English, spoken by 8 female speakers. A sequence of predicted features are produced from a generation algorithm, and the results are randomly degraded at levels from zero to full degradation, for various combinations of the features. Two series of experiments are conducted: the first progressively degrades only single features while leaving all others intact, while the other builds up the system using single, then multiple features. From these experiments, introducing errors into particular articulator-free features, such as vowel, consonant, or sonorant; or articulator-bound features, such as the aspirated feature, pharyngeal features, the nasal feature, the velar feature, or the lateral and rhotic features, do not strongly degrade matching performance. However, matcher performance is more sensitive for errors in the other articulator-free features, and for the articulator-bound features related to vowel place and consonant place, especially, the tongue blade features. For combinations of features, degrading consonantal features, vowel place features, or tongue blade features leads to faster decline in performance, suggesting that these features play more important roles in lexical access.