Modeling per-token variability captures consonant recognition rates that a fixed-rate model misses.

Evidence

The across-listener average data obtained in experiment 1 were averaged across different speech tokens of the same phonetic identity and across the four lowest SNRs (as most confusions occur for low SNRs). The resulting confusion matrices (CMs) are shown as filled gray circles in both panels of Fig. 4 . The model predictions are plotted on top as open black circles. For P (panel a), an underestimation of the recognition for many consonants was observed, indicated by the mismatch of the on-diagonal circles. For Pmod (panel b), a good consonant-specific recognition rate match was found. Both models hit most of the confusions, reflected in the proportion of gray off-diagonal circles matched with a black circle. However, there were also many “false alarms”, particularly in the P-based predictions (panel a). — Dijk, Pim van, Physiology, Psychoacoustics and…, 5616