From Approximation to Validation: Rethinking Frequency-Matched Acoustic Models of Cochlear Implants

This study aims to validate vocoders as acoustic models of cochlear implants by determining whether they capture both perceptual sound quality and speech performance. We hypothesize that valid acoustic models of cochlear implants require listener-specific amounts of frequency mismatch between input filters and output tones or noise bands.

Forty-four adult single-sided deaf cochlear implant users were tested 1 to 5 times for a total of 73 sessions at different time points after initial stimulation. Participants had a cochlear implant in one ear and normal or near-normal hearing in the contralateral ear, allowing within-subject comparisons. In Experiment 1, participants used a method-of-adjustment procedure to select acoustic models most similar to their cochlear implant by adjusting three parameters: low- and high-frequency cutoffs of the acoustic output, and channel interaction (overlap among output sound carriers). In Experiment 2, participants rated the perceptual similarity of five acoustic model types using questionnaires assessing overall similarity and four acoustic dimensions (intelligibility, pleasantness, harshness, and loudness). The five model types included the self-selected model, two all-channel frequency-matched vocoders (tone and noise), and two six-channel frequency-matched vocoders (tone and noise). Frequency-matched acoustic models had output noise bands or tones that were frequency-matched to the analysis filters. In Experiment 3, speech perception was evaluated under six conditions: with the cochlear implant alone and with each of the five acoustic model types presented to the normal hearing ear.

Nearly all participants (69 of 73 sessions) selected acoustic models with frequency ranges different from their clinical frequency allocation tables. The low-frequency edge of self-selected models was significantly higher than clinical allocations (456 Hz for Cochlear Ltd. and 266 Hz for MED-EL). Over 80% of selections used minimal channel interaction (tones, nonoverlapping noise bands, or adjacent noise bands). Self-selected models received significantly higher similarity ratings (mean of 6.11, where 6 means "somewhat similar" and 7 "very similar") compared with all frequency-matched models, which were rated around 3 ("not very similar"). Self-selected models were rated most similar to the cochlear implant across multiple acoustic dimensions and were the only models not rated significantly different from the cochlear implant on any dimension. For speech perception, all-channel frequency-matched models significantly overestimated performance compared with the cochlear implant for both words and sentences. Self-selected models provided speech scores closest to cochlear implant performance. Joint analysis of similarity ratings and speech perception scores demonstrated that self-selected models were the only acoustic models achieving both perceptual similarity to the cochlear implant (rating of 6.11 with 7 being "very similar") and comparable speech perception scores (within five percentage points).

Frequency-matched acoustic models fail to replicate the sound of a cochlear implant and all-channel frequency-matched acoustic models also significantly overestimate speech perception. In contrast, self-selected acoustic models incorporating listener-specific perceptual frequency mismatch provide substantially better matches in both subjective sound quality and speech intelligibility. However, variability in similarity ratings across individuals may suggest that additional perceptual components remain unaccounted for in the current parameter set. These findings question the validity of frequency-matched acoustic models of cochlear implants in research applications.