A log‐linear approach to consonant confusion analysis

Abstract

Log‐linear modeling techniques, along with the likelihood‐ratio chi‐square (G2) statistic, were used to assess the effects of high‐frequency filtering, background noise (shaped to shift thresholds 20 dB), and presentation level on consonant confusion patterns. Specific log‐linear parameters were tested to determine the significance of each of the variables and their associated interactions. Low‐pass filtering changed the pattern of errors relative to unfiltered conditions as presentation level increased. This effect was evidenced by a significant interaction between stimulus, response, level, and filtering. Patterns of errors were also affected by the presence of an external noise. Signal‐to‐noise ratios were selected to produce equivalent performance scores relative to quiet conditions. Thus, even though the same number of errors occurred in the quiet and noise conditions, the relationship between them was significantly different. Interestingly, presentation level was also a significant variable affecting error patterns. Obviously, as presentation level increased, the number of errors decreased. However, these data indicated that the pattern of errors did not remain constant as level changed, and this effect was not similar across filtered or noise conditions. [Work supported by NINCDS.]