Log‐linear modeling of consonant confusion data

Log‐linear models, in conjunction with the G² statistic, were developed and applied to several existing sets of consonant confusion data. Significant interactions of consonant error patterns were found with signal‐to‐noise ratio (S/N), presentation level, vowel context, and low‐pass and high‐pass filtering. These variables also showed significant interactions with error patterns when categorized on the basis of feature classifications. Patterns of errors were significantly altered by S/N for place of articulation (front, middle, back), voicing, frication, and nasality. Low‐pass filtering significantly affected error patterns when categorized by place of articulation, duration, or nasality; whereas, high‐pass filtering only affected voicing and frication error patterns. This paper also demonstrates the utility of log‐linear modeling techniques in applications to confusion matrix analysis: specific effects can be tested; variant cells in a matrix can be isolated with respect to a particular model of interest; diagonal cells can be eliminated from the analysis; and the matrix can be collapsed across levels of variables, with no violation of independence. Finally, log‐linear techniques are suggested for development of parsimonious and predictive models of speech perception.