Cochlear Implant Compression Optimization for Music Listening - Maplaw and AGC
Gilbert, Melanie et al. (2021), Cochlear Implant Compression Optimization for Music Listening - Maplaw and AGC, Dryad, Dataset, https://doi.org/10.7272/Q6JH3JDB
Variations in loudness are a fundamental component of the music listening experience. Cochlear implant (CI) processing, including amplitude compression, and a degraded auditory system may further degrade these loudness cues and decrease the enjoyment of music listening. This study aimed to identify optimal CI sound processor compression settings to improve music sound quality for CI users.
Fourteen adult MED-EL CI recipients participated (Experiment #1: n=17 ears; Experiment #2: n=11 ears) in the study. A software application using a modified Comparison Category Rating (CCR) test method allowed participants to compare and rate the sound quality of various CI compression settings while listening to 25 real-world music clips. The two compression settings studied were: 1) Maplaw, which informs audibility and compression of soft level sounds, and 2) Automatic Gain Control (AGC), which applies compression to loud sounds. For each experiment, one compression setting (Maplaw or AGC) was held at the default while the other was varied according to the values available in the clinical CI programming software. Experiment #1 compared Maplaw settings of 500, 1000 (default), and 2000. Experiment #2 compared AGC settings of 2.5:1, 3:1 (default), and 3.5:1.
In Experiment #1, the group preferred a higher Maplaw setting of 2000 over the default Maplaw setting of 1000 (p = 0.003) for music listening. There was no significant difference in music sound quality between the Maplaw setting of 500 and the default setting (p = 0.278). In Experiment #2, a main effect of AGC setting was found; however, no significant difference in sound quality ratings for pairwise comparisons were found between the experimental settings and the default setting (2.5:1 versus 3:1 at p = 0.546; 3.5:1 versus 3:1 at p = 0.059).
CI users reported improvements in music sound quality with higher than default Maplaw or AGC settings. Thus, participants preferred slightly higher compression for music listening, with results having clinical implications for improving music perception in CI users.
Fourteen MED-EL cochlear implant (CI) recipients (7 men and 7 women, average age 60.3 ± 13.5 SD years) participated in the study. Four subjects were bilaterally implanted and 10 unilaterally implanted, yielding 18 CI ears tested. In Experiment #1, Maplaw manipulations were studied in 17 CI ears and in Experiment #2, AGC manipulations were studied in 11 CI ears.
Thirteen CI recipients were postlingually-deafened and 1 was prelingually-deafened; all used oral/aural communication as their primary method of communication. The mean length of implant use was 3.4 ± 2.3 years and the mean duration of severe-to-profound hearing loss prior to implantation was 8.4 ± 12.9 years (range: <0.1 to 47.2 yrs). The CI users utilized a variety of devices and processing strategies (Table 1). Some of the subjects used a different sound processor model in day-to-day listening than the model used for this research experiment (i.e., a Sonnet), which may have affected results. The Institutional Review Board at the University of California, San Francisco (UCSF) approved this study and informed consent was obtained from all participants.
Nine subjects reported formal music training, beginning at age 8.4 ± 5.4 years and lasting 9.0 ± 7.3 years. Of those nine subjects, three reported being self-taught musicians as did one additional subject who had no formal music training, with training beginning at age 32.0 ± 19.2 years and lasting for 3.9 ± 2.2 years. Six of the CI subjects reported that they do not listen to music in their day-to-day life, while the other 8 subjects listen to music an average of 6.7 ± 4.3 hours per week (range: 2 to 15 hrs/wk).
Most study subjects were using clinical default settings for Maplaw (500, 1000) and AGC (3:1). For the subjects who participated in Experiment #1, three used Maplaw of 500, 12 used Maplaw of 1000-1024, and two used Maplaw of 1500. Of the subjects who participated in Experiment #2, only one was using an AGC setting (3.5:1) in their clinical program different than the default setting (3:1).
A corpus of 25 real-world music clips assembled during earlier studies was used (Roy et al., 2012a, 2012b). Each clip is 5 seconds long; there are 5 clips from each of 5 common musical genres (Classical, Jazz, Country, Rock, and Hip-Hop), and 60% of clips are commonly known while the other 40% are considered obscure. See Roy et al. (2012a) for additional details regarding the selection of musical clips. Music clips were each normalized to 65 dBA. The average acoustic dynamic range for the stimuli was 32.7 dB; by genre, Hip-Hop: 39.9 dB, Jazz: 34.9 dB, Rock: 27.1 dB, Country: 29.0 dB, and Classical: 32.4 dB. For greater detail on the intensity level distribution functions of the music clips, broken down to 1/3 octave bands and grouped by genre, see Figure 1 from Gilbert et al. (2019).
Testing took place in a calibrated sound booth. The stimuli were presented from a speaker located at 0 degrees azimuth with the listener seated 1 meter from the speaker at an average RMS level of 65 dBA.
All CI recipients used a Sonnet sound processor designated for research use only. To increase consistency of audibility of musical clips across test subjects, their preferred everyday fittings were modified as follows: upper stimulation levels (MCLs) were loudness-balanced at 80% of incoming MCLs, and then the new MCLs were swept at 100% to ensure comfort. There were no predetermined restrictions imposed upon the patients’ electrical dynamic range (the difference between MCL and THR). Volume control was fixed at 100%. Microphone directionality and wind noise reduction features were disabled. Other parameters (except Maplaw, AGC, MCLs and THRs) were not altered for either experiment, in order to minimize acclimatization effects to the new programs. For the first 7 subjects, the subjects’ threshold levels (THRs) were measured on each electrode and set at the highest inaudible level. For the remaining 8 subjects, the THRs were set to either 0 or 10% of MCLs. Any changes to THRs were made prior to making other experimental changes.
We then created multiple new programs for the CI user with the Maestro software. For Experiment #1, Maplaw was varied (i.e., 500, 1000, 2000), while AGC was fixed at the default (compression ratio 3:1, sensitivity 75%). For Experiment #2, AGC compression ratio was varied (i.e., 2.5:1, 3:1, 3.5:1), sensitivity was fixed at 100%, and Maplaw was fixed at the default (1000). In the MED-EL signal processing strategies, the sensitivity sets the knee point above which the AGC will become active. At the default sensitivity setting of 75%, the knee point is at 52.7 dB SPL. By changing the sensitivity to 100%, the AGC knee point is changed to 48 dB SPL and so more of the musical stimuli (delivered at 65 dBA) is presented in the dynamic range where the AGC is active.
After the new programs were created, the MCLs were adjusted up or down globally in small (1%) increments in order to obtain equivalent loudness in all programs. This loudness-matching process involved playing a pink noise at 65 dBA and switching back and forth between the default Maplaw program (Program 1) and each of the other two experimental Maplaw programs in turn to ensure that Programs 2 and 3 were loudness-matched to Program 1. Once both Programs 2 and 3 had been individually adjusted to match Program 1, all three programs were cycled through a final time to confirm the pink noise was the same loudness level to the listener in each program. This process was then repeated with the three AGC programs.
Program 1 was always the program with both Maplaw and AGC set to defaults (Maplaw = 1000, AGC = 3:1), while Program 2 and Program 3 were experimental programs and the program order was not randomized. For Experiment #1, Program 2 used Maplaw of 2000, and Program 3 used Maplaw of 500. For Experiment #2, Program 2 used AGC of 3.5:1 and Program 3 used AGC of 2.5:1. It was important to rebalance the loudness within these programs because, without doing so, programs with a higher Maplaw or AGC would seem louder to the CI user, and it is known from previous research that the loudness of music may significantly influence the sound quality perception (Vickers, 2011). For the 10 subjects who completed both Experiments #1 and #2, 8 subjects began with Experiment #1 and 2 subjects began with Experiment #2.
We began the experiments with the intention of including even lower (i.e., more linear) compression settings in the test conditions, but the first few subjects tested with these settings were unable to tolerate these programs. More specifically, when the Maplaw was decreased to 250 or the AGC was decreased to 2:1, the music clips became too soft overall to be evaluated fairly. We attempted to correct for this by increasing the MCLs globally in order to make the music clips equally loud in all of the conditions. Unfortunately, this resulted in the intensity peaks in the music clips becoming uncomfortably loud in the programs with the lowest compression settings and therefore those test conditions (i.e., Maplaw of 250, AGC of 2:1) were discontinued.
CI users listened with the test ear only. If the non-test ear had any residual hearing, a foam earplug was used. In one case, a CI recipient had a Lyric hearing aid in the non-test ear that was situated deep in the ear canal and could not be removed for testing. The Lyric was therefore turned off and functioned as an earplug. All CI users were asked if they could hear the test stimuli with the non-test ear while the earplug was in place and the CI on their test ear was turned off; all indicated they were not able to hear the stimuli with the non-test ear.
Participants completed the experiment using a Dell Latitude E7470. Subjects were given the option of using a touchscreen, mouse, or touchpad. The computer ran the program with MATLAB software version R2012b (Mathworks, Natick, MA, USA).
The method used for testing closely resembles a Comparison Category Rating (CCR) approach, which allows for assessment of relative changes in the perceived sound quality of music across various listening conditions (ITU, 1996). The present study utilized a user-controlled CCR method, which is different from the standard CCR, in that the presentation order is not predetermined or randomized; the users may play the sound samples in any order and with the possibility of unlimited repetitions. Participants were asked to rate the relative sound quality of a particular music clip while using three different programs on their CI (“Prog 1”, “Prog 2”, and “Prog 3”).
The ratings were on a sliding scale with the following categorical labels: “Much better”, “Slightly better”, “Same as Prog 1”, “Slightly worse”, and “Much worse” (see Figure 3). The sliding scale corresponded to numbers from 0 to 200; scores from 0 to 99 represented sound quality worse than Program 1, a score of 100 represented sound quality equal to Program 1, and scores from 101 to 200 represented sound quality better than Program 1. Program 1 was always set to 100, the midpoint (“Same as Prog 1”). The subjects were not explicitly required to use the whole range of the sliding scale. Although the term typically used to describe the results of a CCR-style test is the Comparison Mean Opinion Score (CMOS), per ITU standards (1996), we opted to describe CMOS results using the term “Sound Quality Ratings.”
There were 25 trials in each experiment. Subjects were given verbal instructions, along with a live demonstration of the software interface functionality on the computer, and then were required to complete a practice round supervised by the researcher. The subject was responsible for switching between programs using the FineTuner remote control. The instructions were: (1) Click “Play” to listen to the music clip. You may replay the music clip as often as you wish. (2) Move the Program 1 slider to “Same as Prog 1” (i.e., the midpoint of the scale). (3) Switch the CI processor to Program 2 using the remote control. (4) Play the music clip and move the Program 2 slider to rate the sound quality as it compares to Program 1. (5) Switch the CI processor back to Program 1 and listen to the music clip again. (6) Repeat steps (3) and (4) for Program 3. (7) Click “Save and proceed” when you are satisfied with your ratings of Programs 2 and 3 as compared to Program 1. This procedure enabled a subject to express whether the experimental program(s) provided better or worse sound quality than the clinical default settings, and to what degree.
During the practice round, the subjects were required to complete the task for at least 3 trials (music clips), but were allowed to do additional trials until they reported they were comfortable using the computer, interacting with the software interface, and using the FineTuner remote control. To ensure the programs were changed correctly, the researcher enabled the indicator light on the processor and observed the number of light flashes to verify that they corresponded with the intended program change. During a practice round, subjects used whichever programs they were going to use for the following experiment. Each experiment took an average of 30 to 60 minutes per ear. The CI recipient was given a break between ears (for the bilateral subjects) and between Experiments #1 and #2 (to allow for device reprogramming and to rest).
All statistical analyses were completed using the R statistical computing program (R Core Team, 2020). A linear mixed effects (LME) approach was applied using the “lme4” package (Bates et al., 2015). For specific models and factors, see results below. In order to test for significant differences between test conditions within a single ear, comparisons were made with pairwise t-tests (paired, 2-tailed) and Bonferroni correction was applied.
The Readme-Maplaw.csv and Readme-AGC.csv files describe the data files Maplaw.xlsx and AGC.xlsx, respectively.
The MATLAB package to collect the participant data is included.
The statistical analysis utilized the R statistical program; the R script is included.