The Accuracy of Envelope Following Responses in Predicting Speech Audibility.

Objectives: The present study aimed to (1) evaluate the accuracy of envelope following responses (EFRs) in predicting speech audibility as a function of the statistical indicator used for objective response detection, stimulus phoneme, frequency, and level, and (2) quantify the minimum sensation level (SL; stimulus level above behavioral threshold) needed for detecting EFRs.

Design: In 21 participants with normal hearing, EFRs were elicited by 8 band-limited phonemes in the male-spoken token /susa∫i/ (2.05 sec) presented between 20 and 65 dB SPL in 15 dB increments. Vowels in /susa∫i/ were modified to elicit two EFRs simultaneously by selectively lowering the fundamental frequency (f0) in the first formant (F1) region. The modified vowels elicited one EFR from the low-frequency F1 and another from the mid-frequency second and higher formants (F2+). Fricatives were amplitude-modulated at the average f0. EFRs were extracted from single-channel EEG recorded between the vertex (Cz) and the nape of the neck when /susa∫i/ was presented monaurally for 450 sweeps. The performance of the three statistical indicators, F-test, Hotelling's T, and phase coherence, was compared against behaviorally determined audibility (estimated SL, SL ≥0 dB = audible) using area under the receiver operating characteristics (AUROC) curve, sensitivity (the proportion of audible speech with a detectable EFR [true positive rate]), and specificity (the proportion of inaudible speech with an undetectable EFR [true negative rate]). The influence of stimulus phoneme, frequency, and level on the accuracy of EFRs in predicting speech audibility was assessed by comparing sensitivity, specificity, positive predictive value (PPV; the proportion of detected EFRs elicited by audible stimuli) and negative predictive value (NPV; the proportion of undetected EFRs elicited by inaudible stimuli). The minimum SL needed for detection was evaluated using a linear mixed-effects model with the predictor variables stimulus and EFR detection p value.

Results: of the 3 statistical indicators were similar; however, at the type I error rate of 5%, the sensitivities of Hotelling's T (68.4%) and phase coherence (68.8%) were significantly higher than the F-test (59.5%). In contrast, the specificity of the F-test (97.3%) was significantly higher than the Hotelling's T (88.4%). When analyzed using Hotelling's T as a function of stimulus, fricatives offered higher sensitivity (88.6 to 90.6%) and NPV (57.9 to 76.0%) compared with most vowel stimuli (51.9 to 71.4% and 11.6 to 51.3%, respectively). When analyzed as a function of frequency band (F1, F2+, and fricatives aggregated as low-, mid- and high-frequencies, respectively), high-frequency stimuli offered the highest sensitivity (96.9%) and NPV (88.9%). When analyzed as a function of test level, sensitivity improved with increases in stimulus level (99.4% at 65 dB SPL). The minimum SL for EFR detection ranged between 13.4 and 21.7 dB for F1 stimuli, 7.8 to 12.2 dB for F2+ stimuli, and 2.3 to 3.9 dB for fricative stimuli.

Conclusions: EFR-based inference of speech audibility requires consideration of the statistical indicator used, phoneme, stimulus frequency, and stimulus level.