In contrast to the moderate presentation levels most commonly used in clinical practice, speech encountered in everyday life often occurs at low levels, such as when a conversational partner whispers or speaks from another room. In addition, even when the overall signal level is moderate, levels for particular words or speech sounds, such as voiceless consonants, can be considerably lower. Existing techniques for improving recognition of low-level speech for cochlear implant users include using a wider input dynamic range and elevating electrode threshold levels (T-levels). While these techniques tend to positively impact recognition of soft speech, each has also been associated with drawbacks. Recently, a noise-gating (NG) algorithm was reported, which works by eliminating input to an electrode when signal level in the associated frequency channel is at or below a predetermined threshold. Available evidence suggests that activation of this algorithm can improve recognition of sentences presented at low levels (35 to 50 dB SPL), though it remains unclear whether the benefits will be equally evident with both manufacturer default and individually optimized T-levels. The primary aim of this study was therefore to evaluate the independent and combined effects of NG activation and T-level personalization.
Twenty adults between the ages of 25 and 77 years (M = 54.9 years, SD = 17.56) with postlingually acquired profound hearing loss completed testing for this study. Participants were fit with an Advanced Bionics Naida CI Q90 speech processor, which contained four programs based on each participant’s existing everyday program. The programs varied by the NG algorithm setting (on, off) and T-level method (default 10% of M-level, personalized based on subjective ratings of “very quiet”). All participants completed speech sound detection threshold testing (/m/, /u/, /a/, /i/, /s/, and /∫/), as well as tests of monosyllabic word recognition in quiet (45 and 60 dB SPL), sentence recognition in quiet (45 and 60 dB SPL), and sentence recognition in noise (45-dB SPL speech, +10 dB SNR).
Findings demonstrated that both activating NG and personalizing T-levels in isolation significantly improved detection (speech sounds) and recognition (monosyllables, sentences in quiet, and sentences in noise) of soft speech, with their respective individual effects being comparable. However, the lowest speech sound detection thresholds and the highest speech recognition performance were identified when NG was activated in conjunction with personalized T-levels. Importantly, neither T-level personalization nor NG activation affected recognition of speech presented at 60 dB SPL, which suggests the strategies should not be expected to interfere with recognition of average conversational speech.
Taken together, these data support the clinical recommendation of personalizing T-levels and activating NG to improve the detection and recognition of soft speech. However, future work is needed to evaluate potential limitations of these techniques. Specifically, speech recognition testing should be performed in the presence of diverse noise backgrounds and home-trials should be conducted to determine processing effects on sound quality in realistic environments.