Detectable changes in speech SNR correspond to roughly 26 percent changes in word intelligibility

Evidence

This gave a mean SNR JND of 3.0 dB (σ = 1.0 dB), which, converted to intelligibility using the slope of the psychometric function, corresponded to a mean Intelligibility JND of 26 % (σ= 7.5 %). … Both intervals contained the same randomly chosen sentence, and after both intervals, participants were prompted to respond to which one was clearer. — Dijk, Pim van, Physiology, Psychoacoustics and…, 5673

The JND for intelligibility was estimated by first measuring psychometric functions for IEEE sentences based on keywords at SNRs of ‑16/‑12/‑8/‑4/0 dB or ‑8/‑4/0/+ 4/+ 8 dB. Twenty-four adult participants, median better-ear four frequency average (BE4FA) of 16 dB HL (range ‑3–49 dB HL) and median age of 57 years (range 27–74), were first presented with sentences at ‑16 and 0 dB SNR. Fourteen participants could not respond with half of the keywords (5 keywords/sentence) at 0 dB; those participants were tested at the ‑8–8 dB SNR range. The individual results, averaged across 50 sentences (250 keywords) at each SNR, were fit with a maximum-log-likelihood logistic function. The SNR corresponding to 50 % correct on each individual’s psychometric function (SNR50) was calculated and then used as the reference for an adaptive-track JND experiment; the mean SNR50 was 0.20 dB. This gave a mean SNR JND of 3.0 dB (σ = 1.0 dB), which, converted to intelligibility using the slope of the psychometric function, corresponded to a mean Intelligibility JND of 26 % (σ= 7.5 %). … This raises the question as to whether the JNDs initially measured (in Sect. 2.1 above) were intelligibility rather than SNR JNDs (i.e., were the listeners basing their responses on changes in word intelligibility or in changes in signal-to-noise ratio?) We reasoned that by changing the stimuli to ones with a far steeper psychometric function (MacPherson and Akeroyd 2014 ), namely triple digits, the JND in the appropriate domain would be constant, but the JND in the other domain would change. On testing this hypothesis, we found that the mean SNR JND for digits was 2.5 dB (σ= 0.8 dB) and the mean Intelligibility JND estimate for digits was 17.5 % (σ= 7.2 %). SNR and Intelligibility JNDs were significantly less using digits in unfiltered random (white) noise than sentences in same-spectrum noise [t (23) = 2.98; p = 0.007 and t (23) = 5.00; p = 0.00005]. Across stimuli, both the SNR JND differences (mean 0.5 dB) and the Intelligibility JND differences (mean 8.92 dB) were significantly non-zero (z(23) = 2.98; p = 0.003, and z(23) = 5.00; p ≈ 0). That both differences were non-zero was unexpected; it indicates that one cannot be certain whether the 3 dB SNR JND is indeed the JND for SNR or instead the JND for Intelligibility. — Dijk, Pim van, Physiology, Psychoacoustics and…, 5681

Categorization

• QUESTION: Whoa — a 3 dB change maps to a 26% intelligibility change? That seems enormous, worth checking the slope of the psychometric function and whether this generalizes.
• SEED: Potential essay angle: what listeners actually attend to when asked to detect hearing improvement crosses domains (SNR vs. intelligibility), relevant to hearing-aid benefit claims.
• corroborated: 2 independent passages support this claim — Dijk, Pim van, Physiology, Psychoacoustics and…, 5673; Dijk, Pim van, Physiology, Psychoacoustics and…, 5681