Impact of reduced spectral resolution on temporal-coherence-based source segregation

Abstract

Hearing-impaired listeners struggle to understand speech in noise, even when using cochlear implants (CIs) or hearing aids. Successful listening in noisy environments depends on the brain's ability to organize a mixture of sound sources into distinct perceptual streams (i.e., source segregation). In normal-hearing listeners, temporal coherence of sound fluctuations across frequency channels supports this process by promoting grouping of elements belonging to a single acoustic source. We hypothesized that reduced spectral resolution-a hallmark of both electric/CI (from current spread) and acoustic (from broadened tuning) hearing with sensorineural hearing loss-degrades segregation based on temporal coherence. This is because reduced frequency resolution decreases the likelihood that a single sound source dominates the activity driving any specific channel; concomitantly, it increases the correlation in activity across channels. Consistent with our hypothesis, our physiologically inspired computational model of temporal-coherence-based segregation predicts that CI current spread reduces comodulation masking release (CMR; a correlate of temporal-coherence processing) and speech intelligibility in noise. These predictions are consistent with our online behavioral data with simulated CI listening. Our model also predicts smaller CMR with increasing levels of outer-hair-cell damage. These results suggest that reduced spectral resolution relative to normal hearing impairs temporal-coherence-based segregation and speech-in-noise outcomes.

FIG. 1.

(Color online) Comodulation making release (CMR) stimuli used for computational modeling and behavioral measurements. The stimuli consisted of a target signal (shown in green) in a 100% sinusoidally amplitude-modulated (SAM) tonal complex masker (shown in orange). The masker was composed of three 10 Hz SAM tones at carrier frequencies of 3022 Hz [on-frequency component (OFC)], 2142 Hz (first flanking component), and 4264 Hz (second flanking component). All three SAM tones were presented at the same sound level. In the Comodulated condition, the flanking components were modulated in phase with the OFC, while in the Codeviant condition, they were modulated 180° out of phase with the OFC. The target signal was a 3022 Hz pure tone presented at different signal-to-noise ratios (SNRs). The level of the OFC was fixed while that of the target signal was varied according to the SNR. The total duration of each stimulus was 0.5 s.

FIG. 2.

Across-channel temporal-coherence–based source-segregation model (Viswanathan et al., 2022) to predict CMR (A) and consonant confusions in noise (B) in different vocoding conditions (Intact, Vocoded, Current Spread) and as a function of outer-hair-cell (OHC) damage. (C) The CMR circuit model shown in (A) and (B). Detailed descriptions and model parameters are provided in the main text.

FIG. 3.

(Color online) (A) Model AN-fiber threshold tuning curves, (B), rate-level curves, and (C) thresholds (in dB hearing level, i.e., relative to C_OHC = 1) plotted against ERB [ERB was derived from (A) tuning data, and expressed as a ratio relative to the ERB of a normal-hearing ear (C_OHC = 1)]. Data are shown for varying degrees of OHC damage [simulated by varying the model parameter C_OHC from 1 (normal) to 0 (complete OHC dysfunction)]. All data were obtained for a 3225 Hz CF (i.e., the CF at which we derived CMR predictions).

FIG. 4.

(Color online) CMR as a function of simulated CI vocoding and current spread. (A) Estimated d′ values (mean and standard error across stimulus repetitions) from the temporal-coherence–based source-segregation model [Fig. 2(A)] for different SNRs and CMR conditions (Comodulated, Codeviant). (B) CMR predictions from the temporal-coherence model (mean and standard error across stimulus repetitions), calculated from (A) data as the mean difference in SNR threshold between Codeviant and Comodulated conditions across the d′ values predicted by the model. (C) Behaviorally measured proportion trials correct (mean and standard error across N = 20 samples) for different SNRs and CMR conditions. (D) Behaviorally measured CMR (mean and standard error across N = 20 samples), which was calculated for each sample as the SNR threshold difference between Codeviant and Comodulated conditions at a percent-correct score of 66%.

FIG. 5.

Speech intelligibility in speech-shaped noise as a function of simulated CI vocoding and current spread. The first (leftmost) plot shows predictions from a model of purely within-channel masking (mean and standard error across stimulus repetitions). The second (middle) plot shows predictions from the across-channel temporal-coherence–based source-segregation model [Fig. 2(B); mean and standard error across stimulus repetitions]. The third (rightmost) plot shows behavioral measurements (mean and standard error across N = 48 samples).

FIG. 6.

(Color online) CMR predictions at a fixed OFC level (83 dB SPL) as a function of degree of simulated OHC damage [C_OHC varied from 1 (normal) to 0 (complete OHC dysfunction)]. The sensation level (SL) of the OFC was 63 dB at C_OHC = 1, 43 dB at C_OHC = 0.25, 28 dB at C_OHC = 0.0625, 18 dB at C_OHC = 0.0156, and 18 dB at C_OHC = 0.0039 [derived from Fig. 3(A)]. (A) Estimated d′ values (mean and standard error across stimulus repetitions) from the temporal-coherence–based source-segregation model [Fig. 2(A)] for different SNRs and CMR conditions (Comodulated, Codeviant). (B) CMR predictions from the temporal-coherence model (mean and standard error across stimulus repetitions), calculated as the mean difference in SNR threshold between Codeviant and Comodulated conditions across the d′ values predicted by the model.

FIG. 7.

(Color online) CMR predictions at a fixed OFC loudness as a function of degree of simulated OHC damage [C_OHC varied from 1 (normal) to 0 (complete OHC dysfunction)]. The sound pressure level (SPL) of the OFC was 48 dB at C_OHC = 1, 57 dB at C_OHC = 0.25, 64 dB at C_OHC = 0.0625, 69 dB at C_OHC = 0.0156, and 69 dB at C_OHC = 0.0039. (A) d′ estimates (mean and standard error across stimulus repetitions) from the temporal-coherence–based source-segregation model [Fig. 2(A)] for different SNRs and CMR conditions (Comodulated, Codeviant). (B) CMR predictions from the temporal-coherence model (mean and standard error across stimulus repetitions), calculated as the mean difference in SNR threshold between Codeviant and Comodulated conditions across the d′ values predicted by the model.