The consonant recognition of 17 ears with sensorineural hearing loss is evaluated for 14 consonants /p, t, k, f, s, Ð , b, d, g, v, z, Z, m, n= þ =A=, under four speech-weighted noise conditions (0, 6, 12 dB SNR, quiet). One male and one female talker were chosen for each consonant, resulting in 28 total consonant-vowel test tokens. For a given consonant, tokens by different talkers were observed to systematically differ, in both the robustness to noise and/or the resulting confusion groups. Such within-consonant token differences were observed for over 60% of the tested consonants and all HI ears. Only when HI responses are examined on an individual token basis does one find that the error may be limited to a small subset of tokens with confusion groups that are restricted to fewer than three confusions on average. Averaging different tokens of the same consonant can raise the entropy of a listener's responses (i.e., the size of the confusion group), causing the listener to appear to behave in a less systematic way. Quantifying these token differences provides insight into HI perception of speech under noisy conditions and characterizes each listener's hearing impairment.