(a) Illustration of Probability vs SNR curves for two tokens with the difference in SNR90 values (ΔSNR90) indicated. The SNR90 is defined as the SNR at which the probability of recognition drops to 90%, while ΔSNR90 quantifies the difference in noise-robustness across tokens. (b) The NH ΔSNR90 values for each set of consonant tokens in this study, as computed from NH perceptual data in the presence of speech-weighted noise (Table IV , value for /f/ not shown). These values are computed as in the example of (a) with the male token as talker 1 and the female token as talker 2. For each consonant, a positive NH ΔSNR90 indicates that the female token is more robust to noise, while a negative value indicates that the male token is more robust to noise. The consonants are sorted along the abscissa by NH ΔSNR90. The labels sh = ʃ and zh = ʒ.
Average probability of error (%) over all tested tokens for each HI ear, plotted as a function of SNR [Eq. (2) ] on a log scale. Right ears (R) are shown as solid lines, left ears (L) as dashed lines. The average NH error (gray solid line) is included for reference along with a gray error region representing 1 standard deviation.
Distribution of error for (a) ear 40L and (b) ear 34L at each of the four noise conditions. The abscissa corresponds to the 27 test tokens, sorted for each SNR such that the error increases monotonically; thus, the sort order can vary across ears and SNRs.
Top left and right: Consonant recognition error as a function of SNR, for HI ears (a) 40L and (c) 34L. Each subplot shows the data for one consonant; plots display the error for the female token (diamond marker), male token (square marker), and the average across the two tokens (x marker, dashed line). Bottom left and right: for each consonant [Eq. (3) ], for HI ears (b) 40L and (d) 34L. Consonants are ordered along the abscissa based on the NH ΔSNR90 values (as in Fig. 1 ). is marked for reference. The labels , , and .
(a) for all HI ears. Each point represents the value for a single HI ear, the mean across ears for each consonant is marked with an ‘x’. A negative indicates that the male token has lower error, a positive value indicates that the female token has lower error. Consonants are ordered along the abscissa based on the NH ΔSNR90 values (as in Fig. 1 ). is marked for reference. (b) Comparison and linear regression of the mean values and the NH ΔSNR90 values (see Fig. 1 ), the two values are significantly correlated (ρ = 0.81, p < 0.001). The labels and .
The 17 HI ears are ordered by the average of the left and right ear h 0 values [Eq. (1) ]. The model parameters estimate the flat low-frequency loss h 0 (dB), the frequency at which sloping loss begins f 0 (kHz), and the sloping high-frequency loss s 0 (dB/octave). RMS error ε (dB) of the model fits. The age of the listener and most comfortable level (MCL) for each ear are included. The mean and standard deviation (μ,σ) for all values are reported in the bottom row (ear 14R excluded).
(a) Confusion matrix for the female /bɑ/ token, data from six HI ears (34L/R, 36L/R, 40L/R), at each SNR (dB). (b) Confusion matrix for the male /bɑ/ token, data from the same six HI ears (34L/R, 36L/R, 40L/R), at each SNR (dB). For both confusion matrices, the highest probability confusion in each row is highlighted in bold, and probabilities of 0% are removed to reduce clutter. (c) The recognition data for the female token, averaged across all 17 HI ears; primary confusions are with /d, v, g/. (d) The recognition data for the male token, averaged across all 17 HI ears; primary confusions are with /f, v/. The labels sh=ʃ, zh=ʒ, and a=ɑ.
A confusion matrix showing the average response (%) for each token (average taken over the 17 HI ears and 4 SNRs). Each row contains data for a single token. Confusion probabilities >5% are highlighted in bold, and probabilities <2% are not shown. F, M subscripts denote tokens from female and male talkers.
For each consonant-vowel token (CV), the male (M) and female (F) talker labels are listed, along with the corresponding NH SNR90 values (dB). The /fɑ/ from talker m112 is marked with a * to indicate that this token was not included in the data analysis.
Article metrics loading...
Full text loading...