Stimulation Rate and Voice Pitch Perception in Cochlear Implants

Abstract

The stimulation rate in cochlear implant (CI) sound coding, or the “carrier” rate in pulses per second (pps), is known to influence pitch perception, as well as loudness perception and sound quality. Our main objective was to investigate the effects of reduced carrier rate on the loudness and pitch of coded speech samples. We describe two experiments with 16 Nucleus® CI users, where we controlled modulation characteristics and carrier rate using Spectral and Temporal Enhanced Processing (STEP), a novel experimental multichannel sound coder. We used a fixed set of threshold and comfortable stimulation levels for each subject, obtained from clinical MAPs. In the first experiment, we determined equivalence for voice pitch ranking and voice gender categorization between the Advanced Combination Encoder (ACE), a widely used clinical strategy in Nucleus® recipients, and STEP for fundamental frequencies (F0) 120–250 Hz. In the second experiment, loudness was determined as a function of the input amplitude of speech samples for carrier rates of 1000, 500, and 250 pps per channel. Then, using equally loud sound coder programs, we evaluated the effect of carrier rate on voice pitch perception. Although nearly all subjects could categorize voice gender significantly above chance, pitch ranking varied across subjects. Overall, carrier rate did not substantially affect voice pitch ranking or voice gender categorization: as long as the carrier rate was at least twice the fundamental frequency, or when stimulation pulses for the lowest, 250 pps carrier were aligned to F0 peaks. These results indicate that carrier rates as low as 250 pps per channel are sufficient to support functional voice pitch perception for those CI users sensitive to temporal pitch cues; at least when temporal modulations and pulse timings in the coder output are well controlled by novel strategies such as STEP.

INTRODUCTION

Multichannel cochlear implant (CI) processors have been developed from body-worn devices to behind-the-ear over approximately 40 years of commercial availability. Also, over time, stimulation rates increased with the idea that higher rates per channel should carry more temporal information. Nevertheless, little evidence has been collected supporting this notion. For example, a general finding is that increasing the stimulation rate does not improve speech perception (Vandali et al. 2000; Arora et al. 2009; Brochier et al. 2017) or sensitivity to amplitude modulation (Arora et al. 2011; Green et al. 2012). This may reflect a trade-off between “more detailed” coding and neural effects such as refractoriness and facilitation. At the same time, sound coding methods have changed relatively little, with all systems relying on a bank of bandpass filters to spectrally decompose the input signal for multichannel stimulation. Within each channel, modulations in the signal envelope are then used to control the stimulation intensity of trains of electrical biphasic pulses, usually by varying the pulses’ current amplitude (Envelopes, Fig. 1). The pulse rate of the carrier trains is crucial because it limits the maximum frequency of modulation that can be faithfully represented; the “Nyquist” rate limit applies, such that the carrier rate needs to be at least twice that of the maximum modulation rate that is present in the signal.

Fig. 1.

Temporal envelope modulations captured by windows of different lengths for /ɒ/ vowels with F0 = 164 and 212 Hz: 8 ms windows are used in the clinical ACE strategy and give a good compromise of spectral and temporal resolution with envelope modulations being captured up to F0 = 164 Hz, but poorly at 212 Hz. However, considerable overlap in the base filters produces some spectral smearing between electrodes for the narrowest, low-frequency channels. Three exemplary consecutive electrodes, positioned in the apical-to-basal cochlear direction, are shown in blue, black, and red, respectively. In the Envelopes panel, colored envelopes indicate modulations for low-pitched (F0 = 164 Hz) vowel, while corresponding high-pitched (F0 = 212 Hz) vowel envelopes are always shown in gray for greater clarity; 2 ms allows for minimal 50 % overlap between analysis windows and captures modulation at F0 ≈ 212 Hz and above. 32 ms provides a spectral resolution of harmonics down to F0 = 164 Hz and below: When the 32 ms window analysis is combined with the 2 ms analysis using STEP, the result is minimal spectral smearing with temporal resolution covering up to F0 ≈ 250 Hz. The STEP coding allows a wider range of spectral bandwidths to be specified in the frequency allocation table, which allows for further flexibility and experimentation. In the signal line (top), the blue and red brackets indicate the extent of consecutive sample windows. Synchronizing the 2 ms window to the front of the 32 ms window also produces a shorter latency for STEP, when compared with ACE, when the window length is taken into account (half the window length, 1 ms vs. 4 ms). Modulations are not in phase across channels with ACE but are for STEP (middle row, envelopes vertical gray dotted lines). For reference, the centers for 8 ms and 32 ms bandpass filters are indicated by gray vertical dotted lines in the middle, 2 ms bandpass filter plot. In STEP, the 2 ms bandpass filter (blue) may be used to modulate several channels where it is best aligned with the center of the 32 ms “spectral” band

Carrier rates of 500 pps and 900 pps would faithfully represent voice pitches up to ~ 250 Hz and ~ 450 Hz, corresponding to male voices and to female and children’s voices, respectively. However, most research revealed further temporal constraints for voice pitch perception imposed by cochlear implants. First, CI users are generally poor at detecting modulation at frequencies greater than ~ 150 Hz (Fraser and McKay 2012), irrespective of carrier rate. Another study established upper limits of a temporal pitch at 300 pps and above (Kong et al. 2009), but it was not clear whether loudness cues might contribute to the modulation detection task. Second, they may be unable to rank pitch order even for F0 < 150 Hz correctly (Milczynski et al. 2009), even where modulation depths are greater than those produced in clinical processors. Third, to consistently rank 150 Hz- and 200 Hz-amplitude-modulated pulse trains, a carrier rate had to be above 800 Hz (McKay et al. 1994). Fourth, several methods have been proposed to improve voice pitch perception for CI users by enhancing temporal modulation cues for F0: these include increasing the depth of modulation; synchronizing modulation across channels (Laneau et al. 2006); or imposing a specific modulation shape on, primarily, high rate carrier pulse trains (Green et al. 2005). In the extreme, the carrier pulses are decimated such that the remaining are approximately synchronized to the F0, or, in other words, pulse rate is used to code F0 (Vandali and van Hoesel 2012).

Given these constraints imposed by stimulation rate, we proposed a novel sound coding strategy, called Spectral and Temporal Enhanced Processing (STEP). We designed the STEP to control the spectral and temporal envelopes of the amplitude-modulated pulse trains used to convey information to the auditory nerve in CI users. As in Advanced Combination Encoder (ACE), a widely used clinical strategy in Nucleus® recipients, a single filter bank introduces constraints on the temporal envelope characteristics. The filter bandwidth limits the modulation bandwidth, and there may be phase differences between channels in the modulation produced by a given voice fundamental frequency: this may reduce the salience of the modulation pitch cue for CI users. Conversely, the suppose modulations corresponding to voice F0 are to be faithfully coded, for example, up to 200 Hz or above. In that case, this places limitations on the spectral resolution with a single bank of filters. Thus, we designed the STEP coder to overcome these limitations while retaining ecological signal-driven information about the sound source rather than imposing synthetic modulation waveforms on the output signal.

Two primary objectives of this study were to compare voice pitch between the ACE and the STEP and to investigate the effects of reduced carrier rate using STEP coding on voice pitch perception. In Experiment 1, we compared STEP with the clinical stimulation strategy, the ACE, developed by Cochlear©, using subjects’ clinical carrier rates (720–1200 pps). Loudness rating as a function of speech intensity was used to establish equivalent loudness between programs for the pitch perception tests that followed. Temporal pitch ranking was evaluated using vowel-like stimuli with F0 range 126–212 Hz, and functional voice pitch perception was evaluated using a 40-talker voice gender categorization test (Kovačić and Balaban 2009). By carefully controlling the coded speech signals’ modulation characteristics using STEP coding, it was then possible in Experiment 2 to investigate the effects of three carrier rates (250, 500, and 1000 pps) on loudness perception, pitch ranking, and voice gender categorization. Pilot testing indicated, as expected, that using a constant rate 250 pps, carrier resulted in pitch reversals, and thus a novel algorithm 3Quick, which synchronized stimulating pulses to envelope peaks, was used to produce nominal 250 pps carriers (Q250 condition). We hypothesized that reducing the stimulation rate even to 250 pps per channel is sufficient to support functional voice pitch perception for those CI users sensitive to temporal pitch cues.

METHODS

Spectral and Temporal Enhanced Processing (STEP) Coding

In the current study, we used the STEP coder with a dual filter-bank signal analysis system, such that a spectral profile filter bank (SP) was used to control the stimulation levels applied to different electrodes, and a parallel temporal envelope filter bank (TE) was used to enhance amplitude modulation across electrodes (Fig. 2). Signal analysis was performed at ~ 1000 pps so that modulation frequencies up to ~ 500 Hz were sufficiently sampled. A first-order integrator provided control of the frequency range of the modulation signal applied to each channel separately. The integrator provided the first-line method to overcome Nyquist aliasing of modulation frequencies higher than half the envelope sampling rate. For example, for a 500 pps carrier, every other sample was used after the integrator had been applied. Further, for low carrier rates, a peak-synchronizing algorithm was implemented to place carrier pulses at the time of peaks in the envelope signals to circumvent the Nyquist rate limitation (Fig. 3a). The aim was to maintain temporal pitch perception over the range of F0s to be evaluated.

Fig. 2.

Flow diagram for ACE and STEP sound coding. Signal pre-emphasis and gain are applied at the input. ACE coding used the signal path only as indicated by the dashed lines in the lower part of the figure, and thus a single filter bank based on 128 pt (8 ms) window. The STEP coding used both the upper and lower paths: a short 32 pt (2 ms) window was used to generate modulation envelopes in the temporal envelope path. A long, 512 pt (32 ms) window was used in the spectral profile path to produce high-quality basis filters spaced at 62 Hz. For both STEP and ACE, the basis filters (bins) are combined to form up to 22 spectral bands, defined by the frequency to electrode allocation table. In STEP, the upper path’s envelopes are used to modulate the spectral estimates from the lower path. Then, for both STEP and ACE, the filter bank’s channel outputs are sampled at the stimulation rate, and the N largest selected for stimulation. The envelope amplitudes are then mapped to the threshold and current levels for each electrode. Synchronization may be applied for STEP coding such that peaks in the temporal envelope signal in the upper path define the timing of stimulation pulses. The synchronization strategy (3Quick) adds intermediate pulses at the nominal rate (250 pps in these experiments, i.e., every 4 ms) after peak pulses (Q250, Fig. 3a)

Fig. 3.

The effect of carrier rate on the temporal representation of F0: a Pulse sequences produced by the STEP coder for rates 250 to 1000 pps; including the synchronized modulation condition Q250, for a vowel with F0 = 164 Hz. Where the Nyquist condition is not met, the inter-peak interval does not correspond to the F0 period (T0). b F0 estimates based on temporal peak intervals for eleven vowel stimuli with actual F0 126 to 212 Hz. STEP produced more accurate temporal modulation than ACE for frequencies greater than ~ 160 Hz. Where the Nyquist criterion was not met, as in the pure 250 pps condition, aliasing occurred. However, this was remedied using the synchronized Q250 condition. c Pilot test for vowel pitch ranking by subject CI16 gave expected results in that pitch ranking was poor for the pure 250 pps condition, compared with others. Also, there was a slight reduction in performance using ACE versus STEP for a 1000 pps carrier

Subjects

We recruited sixteen subjects implanted with Nucleus devices manufactured by Cochlear Corporation (seven females and nine males; age range: 13–46 years, mean 21.9 years, SD 10.2 years), with at least 3 years of CI experience. Demographic characteristics are given in Table 1. The subjects were pooled from a more extensive CI subjects database registered at the Polyclinic SUVAG, Zagreb, Croatia, which provides health and audiological care, and speech rehabilitation for hard of hearing and language-impaired patients. As part of that care, participants underwent non-verbal psychological assessments by professional psychologists, which showed that none of them had reduced cognitive abilities that might compromise performance. This study’s data collection was approved by the ethical committee of the SUVAG Polyclinic (# 510–08/17–20). Subject CI16 is also the first author.

Table 1.

Demographic and CI characteristics of the subject population

Subject	Age (years)	Age at implant (years)	Sex	Ear	Implant type	Strategy	Clinical stimulation rate (pps)	Number of active electrodes	Number of maxima	Experiments
CI1	46	38	F	R	CI24RE(ST)	ACE	900	21	8	1,2
CI2	22	7	F	L	CI24M	ACE	900	20	8	1
CI3	28	13	M	R	CI24R (CS)	ACE	720	20	8	1
CI4	22	9	F	R	CI24R (CS)	ACE	900	17	8	1
CI5	25	11	M	L	CI24R (CS)	ACE	900	20	8	1
CI6	21	2	F	R	CI24M	ACE	1200	20	10	1,2
CI7	21	7	M	R	CI24R (CS)	ACE	1200	19	8	1
CI8	15	3	F	R	CI24R (CA)	ACE	900	21	10	1,2*
CI9	15	3	M	R	CI24R (CA)	ACE	900	21	8	1,2*
CI10	16	5	M	L	CI24R (ST)	ACE	1200	22	8	1,2
CI11	13	9	F	L	CI24RE (CA)	ACE	900	22	8	1,2
CI12	13	2	M	R	CI24R (ST)	ACE	1200	22	8	1,2
CI13	14	3	F	R	CI24R (ST)	ACE	1200	21	6	1,2¥
CI14	18	2	M	R	CI24R (CS)	ACE	1200	21	8	1,2
CI15	16	4	M	R	CI24R (CA)	ACE	1200	20	8	1,2*
CI16	45	29	M	R	CI24R (CS)	ACE	1200	17	8	2

*These subjects only completed the loudness rating and voice gender categorization in Experiment 2 due to poor pitch ranking in Experiment 1

^¥Subject did not finish pitch ranking or voice gender categorization for Q250 due to an error in coding

Sound Coding

Sound coding for Nucleus cochlear implants was performed using a benchtop system, where pre-recorded speech stimuli were processed in real-time. The benchtop system allowed emulation of the clinical ACE strategy (Vandali et al. 2005) as well as the novel strategy STEP, which we describe here. Both strategies consist of four basic stages: windowing and fast Fourier transform analysis (FFT) of the input signal, combining the frequency bins into spectral bands to allow envelope extraction, applying the gain, and modulating the carrier pulse trains (Fig. 3). For both strategies, spectral maxima selection (Vandali et al. 2005) is performed, such that a subset of the total number of channels is selected on each stimulation cycle; most subjects used eight to ten maxima in their clinical processors (Table 1).

ACE Filter Bank

Speech stimuli were sampled at 16 kHz. Prior to spectral analysis, pre-emphasis was applied to the input signal, and a low pass filter applied at 4 kHz to mimic the frequency response of the standard directional microphones used in clinical processors. For the ACE strategy, a single bank of phase quadrature envelope filters was implemented using an 8 ms Hann window and 128-point FFT (Fig. 2, processing blocks joined by dashed lines). This produced bandpass filter bins with a minimum bandwidth of ~ 250 Hz, spaced every 125 Hz (Fig. 1, left column). These filter bins were then combined using vector summation to produce the necessary number of bandpass filters as specified in the frequency-to-electrode allocation table (FAT). The combined bins approximate the phase quadrature response of finite impulse response filters with similar time window characteristics (i.e., 8 ms). Here, and for the STEP processing described below, the sensitivity of bandpass filters was equalized by applying fixed gains. We note that no additional processing, such as automatic gain control or noise reduction, such as found in clinical processors, was applied to either ACE or STEP programs.

STEP Filter Banks

After input signal filtering, two spectral analysis paths, TE and SP, are applied in the STEP strategy. In the TE path, a short, 2 ms Hann window is applied on each analysis cycle, followed by a 32-point FFT (the upper path in Fig. 3). This produces bandpass filter bins ~ 1000 Hz wide, spaced at 500 Hz to extract temporal envelope features, including modulations at F0 (Fig. 1, center column). In the SP path, a longer Hann window is applied: Each 32 ms window of signal data is cut in half and overlapped and added, then a 256-point FFT is applied (Fig. 2, all lower paths). This produces bandpass filter bins 62 Hz wide, every 62 Hz. The envelope amplitude from a group of these bins is summed together according to the FAT. The resulting filters are of a higher quality than those in the ACE strategy so as to produce much less overlap between channels (Fig. 1, right column, Band-pass filters, and Electrodes). The long SP window’s temporal response obviates the need for averaging between analysis cycles to obtain stable spectral energy estimates (Fig. 1, right column, Envelopes).

The TE path’s output amplitudes are greater than those produced by the SP path simply due to the width of the TE filters; the wider the filter, the more energy is captured. Therefore, an additional estimate of the average power of TE channels is made using the second grouping of SP bin amplitudes, which reflect the frequency response of the TE filters (Fig. 2, middle, TE energy corrections). Final output amplitudes are essentially the SP path amplitudes multiplied by the normalized TE amplitudes.

Further TE Path Processing

The output amplitudes from the TE path (a_TE) are further processed: firstly, a simple integrator is applied to each channel, which acts as a variable low-pass filter. The integrator time constant was controlled using the smoothing factor α, such that the envelope amplitude after integration a_i = (1-α)·a_i-1 + α·a_TE. Theoretically, α = 0.5 gives a time constant equal to the envelope sampling period (1 ms): In pilot testing by the first author, α = 0.5 gave both good sound quality and optimal pitch ranking (see below): Using α > 0.5 produced poorer sound quality, and second harmonics of F0 were plainly visible in the TE outputs. The sound quality was also impaired for α <  < 0.5. No integration (α = 1.0), as for ACE, was used in Experiment 1 and, α = 0.5 in Experiment 2.

The more precise timing of peaks in the integrated temporal envelopes was estimated by parabolic interpolation between the peak envelope sample and adjacent samples (see the Q250 condition below).

Amplitude Mapping

Amplitudes at the output of the ACE and STEP filter banks were converted to current levels using the system conventionally used in Nucleus implants (James et al. 2003): An instantaneous input range window of 40 dB was mapped to the electric dynamic range defined by subjects’ clinical threshold (T) and comfort (C) levels obtained with stimulation rates between 720 and 1200 pps (Table 1). Nominal input sensitivity was set to give approximately 2 % saturation of the mapping function using speech signal levels of − 25.0 dB RMS using the ACE strategy. The input gain to the system was adjusted to obtain equal loudness across strategies and stimulation rates, as discussed below.

Analysis Cycles and Pulse Sequencing

In Experiment 1, the stimulation rate was at the subject’s usual MAP rate for both ACE and STEP strategies. The clinical stimulation rates of 720, 900, and 1200 pps were approximated by 727, 889, and 1231 pps. Signal analysis was performed at the stimulation rate for both strategies. The channels with the highest amplitudes were selected for stimulation in each cycle, and electrodes stimulated in the base-to-apex order, with equal time intervals between each channel. Channels with amplitudes below the input range window were not stimulated.

In Experiment 2, signals were analyzed at 1000 frames per second for all conditions, giving an optimal 50 % overlap between TE signal windows (16 samples). Integration was then applied to channel envelopes. For the 1000 pps condition, all envelope samples were used, in the 500 pps every other sample, in the 250 pps, fixed-rate, every fourth. Based on modulation peak intervals (Fig. 3a), we may estimate the perceived pitch order of vowels. Higher carrier rates are likely to be more accurate in representing particular F0 periods; however, the perceived pitch is likely to be based on some average of such peak intervals (Carlyon et al. 2002). A reasonable estimate of F0 should be obtained in this way, at least for longer vowels, and where the Nyquist requirement is met. We ran the eleven vowel stimuli used in our experiment through the STEP coder. Figure 3b gives the estimated F0, based on averaged inter-peak intervals, versus the actual F0. It can be seen that the estimated F0 is close to the actual F0 for STEP using 1000 and 500 pps. However, for STEP running at 250 pps, aliasing is evidently such that the estimated F0 decreases with increasing actual F0 due to pitch peaks being missed by the pulse sequence (Fig. 3a). The Q250 processing is similar to the FAST strategy (Smith et al. 2014). In the Q250 processing condition, the carrier pulses were synchronized to the peak timings. Unlike FAST, each peak synchronized pulse was then followed by regular-timed pulses at nominally 250 pps until the next peak was detected (Fig. 3a, Q250). The estimated F0 for Q250 was found to be approximately equal to the actual F0 up to 190 Hz (Fig. 3b).

Running the vowels through ACE resulted in accurate F0 estimates in the lower range, with poorer estimates for F0 > 164 Hz. This may be explained by low modulation depth being produced in some channels compared with others (Fig. 1, left column, F0 212 Hz in gray versus 164 Hz). Our F0 estimation did not consider the potential effects of loss of synchronized modulation between channels (Laneau et al. 2006), which could reduce modulation cues due to current spread and auditory neurons “receiving” current from several channels at a time.

The sound coder conditions were pilot tested by the first author (Fig. 3c). Vowel pitch ranking was satisfactory for 1000 and 500 pps carriers, but irregular and poor overall across the F0 range for 250 pps, as predicted by the Nyquist requirement and the F0 estimates described above (Fig. 3b). However, the synchronized Q250 condition gave pitch-ranking functions similar to those where the Nyquist requirement was met, and therefore, this was used as the low-rate condition in Experiment 2.

Experiment 1 ACE versus STEP Coding

In Experiment 1, ACE and STEP coder conditions were compared using clinical rates and maxima. Loudness ratings, pitch ranking, and voice gender categorization were measured for subjects 1–15 in Table 1, according to the procedures described below. The loudness rating procedure allowed the setting of input gain to the coders to achieve subjectively equal loudness for pitch and gender tests.

Experiment 2 STEP Coding at Three Different Stimulation Rates

In Experiment 2, we compared STEP coding at three different rates (1000 pps, 500 pps, and Q250 pps) using a clinically determined number of maxima. Loudness ratings were used to set input gain to obtain equal loudness between rate conditions. Eleven subjects completed gender testing. However, only those eight subjects achieving d-prime > 1 on pitch ranking for either coder or F0 range in Experiment 1 performed pitch ranking tests in Experiment 2.

Procedures

The experimental apparatus consisted of an L34 cochlear implant research processor (Cochlear Corporation), which was connected to a computer controlling stimulus delivery by custom-written software. For each subject, we first identified their everyday clinical map using standard CI fitting software (CustomSound V4.1) whose parameters (coding strategy, type of stimulation, T, C levels for each electrode, frequency allocation table for electrodes, stimulation rate, number of maxima, interphase gap, jitter) were read-out and transferred to the stimulus coding and delivery program. The program was written in Python 2.7 (Python Software Foundation, https://www.python.org/) and used the NIC 3 interface (Cochlear Corporation) to deliver electric pulses to the implant via the L34 processor. Technical verification of the experimental apparatus running ACE in pilot tests prior to this study showed neither perceptual nor electrodogram differences between the clinical processor and the experimental apparatus used in this study.

Both experiments involved loudness rating, vowel pitch ranking, and voice gender categorization following the same procedures described below. Each subject was given instructions before the start of the experiment and had practice trials to ensure a correct understanding of the task. We note that MAP T and C levels were fixed in each experiment (i.e., the same for all rates in Experiment 2).

Loudness Rating

Subjects estimated the stimulus loudness by rating on a number scale from 1 to 7, indicating the softest and the loudest possible sound, respectively. To ensure that the participants understood the task, they were first presented with the graphical scale representing loudness. We used two speech stimuli tokens, one male with F0 = 102 Hz and a female with F0 = 184 Hz, taken from the speech corpus developed for voice gender categorization (Kovačić and Balaban 2009, also see below the section of Voice gender categorization). The stimuli were presented at seven intensity levels chosen in random order in the attenuation range between − 20 dB and − 30 dB rel. to C level for ACE and in the range − 25 dB to − 35 dB rel. to C level for STEP. Although the ACE and STEP filter banks’ nominal sensitivity for pure tones was equalized, the width of the SP filters and the combination with TE filters for STEP produces generally lower output amplitudes than for ACE for complex signals. We estimated that STEP produces amplitudes approximately 5 dB softer than ACE and using the different amplitude ranges would produce comparable perceptual loudness ranges for the two coders. Each stimulus was presented twice at each intensity level. To account for a potential variability between responses to the same stimulus intensities, we measured response consistency as an absolute difference between the two responses. In Experiment 1, which compared ACE vs. STEP stimulation, loudness was estimated both for ACE and STEP using the clinical stimulation rate, while in Experiment 2, loudness was estimated for the three rate conditions of Q250, 500 pps, and 1000 pps, separately. Importantly, we used this loudness rating data to match comparable loudness between coding strategies and between different stimulation rates for subsequent pitch-ranking tests.

Vowel Pitch Ranking

Subjects ranked pairs of bandpass filtered (300–1200 Hz) /hɑd/ vowels (Massida et al. 2013) with F0 ranging from 126 to 212 Hz (Fig. 3b). Stimuli were 300 ms long with a 10 ms linear ramp; the inter-stimulus interval was 200 ms. The eleven stimuli were grouped into two blocks with reference F0s 126 and 164 Hz, and 164 and 212 Hz. Intermediate comparison frequencies were 134, 141, 147, and 155 Hz for low F0 range, and 175, 181, 191, and 204 Hz for high F0 range. There were six trials per pair, making a total of 54 trials per block. In each trial, one reference stimulus of the pair was first presented twice, followed by the other (3 trials of each order). Subjects were asked to indicate whether the last stimulus was higher or lower in pitch than the first two stimuli. Stimulus trials were presented in random order within a block. In Experiment 1, all stimuli were presented at the stimulus level according to the determined individual gain matches for ACE and STEP. The stimulation level corresponded to a subjective loudness of 4.5 units on a scale from 1 to 7, determined during the loudness rating. Before the start of Experiment 1, the experimenter explained the pitch-ranking task with a graphical sketch of low and high pitches, and when necessary, allowed the subject to perform up to 6 practice trials to assure they understood the task.

Loudness Discrimination and Roving

In Experiment 2, coded stimuli for vowel pitch ranking were also balanced overall for loudness; however, additional loudness roving was added to reduce the potential effects of loudness cues in pitch ranking at different rates (Galvin and Fu 2009; Fraser and McKay 2012; Galvin et al. 2014). Subjects ranked loudness using all eleven vowel stimuli. Trials consisted of the same vowel presented twice at the same level, followed once at a level either 2, 4, or 6 dB softer or louder. All vowel stimuli were presented with each intensity difference, making a total of 33 trials, presented in random order. The lowest intensity difference which produced at least 8/11 correct responses for the 500 pps condition, was used as the roving level for all pitch ranking in Experiment 2. For each pair of reference-comparison F0s, two trials contained an increase in intensity for the last interval, two a decrease, and two no difference.

Voice Gender Categorization

Subjects were assessed for voice gender categorization with the 2-alternative forced-choice task using the same stimulus set of voices from Kovačić and Balaban 2009 study. All speech items in the stimulus set were two-second-long passages from news announcements spoken by 20 female and 20 male speakers, who were professional speakers of the national broadcasting company; stimuli were characterized by the mean fundamental frequency (F0) of the utterance. These 40 voices formed a continuum starting from the lowest male speaker with F0 = 81.8 Hz and ending with the highest female speaker with F0 = 218.7 Hz. Mean F0s were 183.3 ± 5.4 Hz and 117.9 ± 4.8 Hz across female and male voices, respectively (see Fig. 1 in Kovačić and Balaban 2009). We note that vocal tract length (Smith et al. 2007) also varied in these stimuli, particularly between male and female speakers, but this had not been found to be systematically used by CI users in previous studies of the stimuli (Kovačić and Balaban 2009). After listening to a single speech item spoken by either a male or female speaker, the subject was requested to categorize the gender of the speaker by pressing the appropriate button representing a male and a female speaker. There was no feedback nor the possibility to replay the trial. Before the commencement of the task, the experimenter explained the procedure to the subject with the graphical sketch in order to be assured that the subject understood the task. The task with a total of 40 trials usually lasted between 5 and 8 min. This procedure for voice gender categorization was the same for both Experiments 1 and 2. None of the speech tokens used for voice gender categorization were taken for loudness rating to avoid familiarization.

Data Analyses

In order to obtain gain settings for equal loudness between coders and rates, linear functions were fitted to individual subjects’ data for each condition. Where possible, gain settings giving a loudness rating of 4.5 were used in each experiment. However, lower rate conditions could require a gain setting which would produce considerable saturation at C level in some subjects; in these cases, the equal-loudness level was reduced such that the highest gain applied was − 15 dB, and gains for that loudness were determined for the other conditions.

Loudness ratings were analyzed using linear mixed models (LMM) to obtain generic functions of loudness rating versus intensity across subjects, with stimulus gender and coder condition as factors for both experiments. Two random-effects models were used; one which assumed a constant slope of loudness rating versus intensity, and a second where the slope varied by subject. Models were checked for general validity using quantile–quantile plots and other diagnostics. We performed LMM within the R environment, version 3.6.3 (R Core Team 2020); we specifically used the function “glm” from the core package “stats”.

Pitch-ranking functions were plotted for each coder and rate condition, and each reference F0 and F0 range (e.g., Fig. 3c). This allowed determination of the normality of psychometric functions with increased performance from a chance to perfect score with increasing F0 difference, as well as the presence of pitch reversals and non-monotonic functions. Data per F0 range and condition were reduced to “global” d-prime scores, using hit rates for higher/lower across all reference/comparison F0s. For Experiment 1, paired t-tests were used to compare ACE with STEP. For Experiment 2, linear mixed models were used to compare 1000 pps, 500 pps, and Q250 coder conditions.

Scores on the voice gender categorization test were analyzed using paired t-tests and linear mixed models for Experiments 1 and 2, respectively. Also, sigmoid discrimination functions were fitted using logistic regression for each coder condition in Experiment 2 to obtain an aggregate voice gender category stimulus F0 functions.

RESULTS

Experiment 1

Loudness Rating

A repeated-measures analysis based on linear mixed models revealed significant effects of speaker gender (χ²[1] = 75.8, p < 0.001) and coding (χ²[1] = 254.1, p < 0.001) on loudness rating. In addition, we found that loudness rating increased from 0.16 to 0.34 rating for each dB gain in the input amplitude (χ²[1] = 155.2, p < 0.001; varying by subject χ²[2] = 17.0, p < 0.001). Adding interactions between factors did not significantly improve the LMM (χ²[3] = 4.6, p = 0.20), indicating that loudness-vs-amplitude slopes were similar for gender and coder type. The mean slope was 0.24 loudness rating units per dB. On average, STEP was 1.54 loudness rating units softer than ACE, equating to 6.8 dB. The male speaker stimulus was 0.68 loudness rating units softer than the female stimulus, equating to 3.0 dB. Variability in the responses to the same stimulus level was small; the average consistency of all subjects in the ACE was 1.00 ± 0.08 units of loudness rating, while for STEP was 0.93 ± 0.08 units of loudness rating.

We further investigated the difference in loudness found between the female and male voice stimuli: the stimulus waveforms were equalized in terms of RMS amplitude prior to coding. However, pre-emphasis filtering was applied before the filter bank (Fig. 2). The RMS amplitude after pre-emphasis was 3 dB higher for the female voice stimulus than for the male voice stimulus, thus explaining the loudness difference and implied the loudness rating procedure was valid. Figure 4a and c show distributions of pulses across intensities for the male and female voice, while Fig. 4b and d show distributions of the total energy per channel for male and female voice. The overall difference in coder output by voice gender is also evident in the center of gravity (CoG) of the amplitude histogram (points, Fig. 4a). The CoGs for STEP and ACE were well aligned using the subjective loudness estimates.

Fig. 4.

Statistics of output pulse characteristics compared to ACE and STEP processing. The two speech stimuli, one male and one female, used for loudness scaling, were processed with coders at 1000 pps. Attenuation levels for ACE and STEP were set according to the averaged attenuations of all 15 subjects producing a perceptual loudness rating of 4.5 (− 28.4 dB for ACE coding and − 21.7 dB for STEP, a difference of 6.7 dB). a and c Total number of electrical pulses distributed by intensity level for male and female speech item, respectively. Centers of gravity were calculated for each distribution (CoG); b and d Total energy (defined as the cumulative sum of pulse amplitudes) for each channel for the male and female speech items, respectively. Channel 1 corresponds to the lowest frequency electrode, while 22 to the highest, contrary to the usual Cochlear Nucleus electrode numbering. All total energy values were normalized to the total sum of all cumulative sums of pulse amplitudes of the STEP female item

For the female voice in Fig. 4b, but also to some extent the male speaker, more stimulation was seen in the high-frequency channels for STEP (12–21) than the low (1–11) compared with ACE. This is mostly a result of the low-frequency spectral filters in STEP being significantly narrower than for ACE (Fig. 1, bandpass filters 32 ms vs. 8 ms), and thus the maxima selection more often selected high-frequency filters. Bands 11 and above had much more similar widths.

Vowel Pitch Ranking

Individual global d-primes for pitch ranking are compared for ACE and STEP in Fig. 5a and b, for lower and higher F0 ranges. Subjects exhibited a range of vowel pitch ranking abilities, with 9/15 subjects performing at or near chance even in their best condition. D-primes greater than ~ 1 were generally associated with the expected “psychometric” pitch-ranking functions, while increasing differences in F0 between the reference and comparison frequencies were generally associated with improved ranking performance (e.g., Fig. 3c). For the low F0 range, global d-primes for pitch ranking were not significantly different for STEP versus ACE (t[14] = 0.166, p = 0.87). However, mean d-primes were significantly greater, indicating better performance, for ACE than for STEP for the high F0 range (difference 0.42, t[14] = 2.397, p = 0.031). Two notable outliers from equal performance using the two coding methods were seen in the high F0 range: One case (CI8) exhibited pitch reversals consistently for both reference frequencies as if aliasing were occurring (as described in the Methods: Sound Coding section for a low-rate carrier, e.g., Fig. 3a). In another (CI7), good monotonic pitch-ranking functions were obtained with ACE, but a random pattern was obtained using STEP. For CI8, STEP appeared to facilitate pitch ranking in the low F0 range (Fig. 5a, inset). For CI7, their otherwise excellent performance, and therefore sensitivity to temporal pitch, for both coding methods in the low F0 range suggests that there may have been some other reason for poor performance using STEP in the high F0 range. We note that CI7’s pitch-ranking functions followed a monotonic pattern with good performance using both 500 pps and 1000 pps rates in Experiment 2. This implies that the higher modulation bandwidth used in Experiment 1 using STEP with α = 1.0, that would include harmonics of F0, may have been the cause of the pitch confusion. Removing CI8, with the pitch reversals in the high F0 range, rendered the ACE-STEP difference non-significant (t[13] = 2.028, p = 0.64).

Fig. 5.

Pitch ranking results from Experiment 1. a The group average global d-prime scores for STEP versus ACE for the low (left) and high F0 ranges (right). Whiskers indicate standard deviations; b Global d-prime scores for STEP plotted against those for ACE, for the low (left) and high (right) F0 ranges. Insets are the results for one subject CI8, who exhibited the expected near-monotonic pitch ranking functions using STEP in the low F0 range but had pitch reversals in the high F0 range. Short diagonal lines on the line of equivalence indicate the threshold for performance significantly above chance, or global d' > 1

Voice Gender Categorization

The performance score in the voice gender categorization task was measured as the proportion of correct responses. All but one subject could categorize gender at a level significantly above chance (> 26/40). Mean scores for ACE and STEP were equal (34.8/40, t[14] = 1). There was no apparent bias, indicated by no difference in mean scores by gender.

Experiment 2

Loudness Rating

Loudness ratings are plotted in Fig. 6a for female and in Fig. 6b for male speaker stimuli. A repeated-measures analysis based on linear mixed models revealed significant effects of gender (χ2[1] = 73.2, p < 0.001) and rate (χ2[2] = 84.7, p < 0.001), as well as a significant slope of loudness rating versus input amplitude (χ2[1] = 178.3, p < 0.001) which varied between subjects (χ2[2] = 54.4, p < 0.001): from 0.13 to 0.24 rating units per dB. Adding an interaction between input amplitude and rate significantly improved the LMM (χ2[2] = 6.58, p < 0.05): the interaction effect was found to be produced by the 1000 pps condition compared with the other two conditions (t[369] = 2.40, p < 0.05).

Fig. 6.

Loudness rating versus input level at three rates for Experiment 2, shown separately for female (panel a) and male stimuli (panel b). Box plots and raw data are given for loudness ratings at each input level. Lines indicate the results of mixed linear modeling of loudness ratings for each rate. Ratings for 1000 pps were significantly higher than for both 500 and Q250, and there was a significant difference in ratings between female and male speaker stimuli. Box midpoints are medians, upper and lower box edges 25/75 percentiles, and whiskers indicate 5/95 percentiles

Slopes of loudness versus amplitude were slightly steeper for 1000 pps; on average, 0.19 rating units per dB, compared with 0.17 for 500 and 250 pps. However, most of the difference was due to ratings for the highest input levels. In summary, a linear fit of loudness rating versus input level in dB produced satisfactory model properties excepting some mild deviation at the lowest end of the range for all rates, as well as the highest end for 1000 pps. In the 1000 pps condition, it was 1.25 rating units louder than 500 pps (t[359] = 4.38, p < 0.001). However, Q250 gave loudness approximately equal to that for 500 pps (t[359] =  − 0.212, p = 0.83). The difference in loudness between 1000 and 500 pps equated to 3.5 dB on average (2.4–4.4 dB across subjects). The male speaker stimulus was 0.51 rating units softer than the female stimulus, equating to 3.0 dB; the same difference was found in Experiment 1.

Vowel Pitch Ranking

Global d-primes for pitch ranking were compared for different rate conditions in Fig. 7a and b, for low F0 and high F0 ranges, respectively. Overall, the rate had little or no effect on d' for the low F0 range, with the few scores d' > 1 not being systematically affected by changes in rate condition from Q250 to 1000 pps. However, for the high F0 range, d' seem to increase with the rate; statistically, for 1000 pps d' was significantly higher compared with the Q250 condition (t[13.2] =  − 2.82, p < 0.01), but not compared with the 500 pps condition (t[13.1] =  − 1.67, p = 0.12). Interestingly, in three subjects, global d' for pitch ranking was significantly reduced in the high F0 range for the Q250 condition, falling to chance performance and even inducing pitch reversal for CI11. C11’s pitch ranking was consistently good (global d' > 1) across all rate conditions for the low F0 range; we may interpret this as their modulation pitch limit being close to the low/high F0 boundary and perhaps therefore very sensitive to carrier rate.

Fig. 7.

Pitch ranking and gender identification results from Experiment 2. a Global d-prime scores for the low F0 range and b, high F0 range; c Voice gender categorization score by stimulation rate condition. Error bars indicate standard deviations

Voice Gender Categorization

Voice gender categorization scores are compared for different rate conditions (Fig. 7c). Scores were significantly greater for 1000 pps (least-square mean score, LSM 37.2) compared with Q250 (LSM 35.6, t[19.5] =  − 2.27, p < 0.05) and 500 pps (LSM 35.6, t[19.2] = -2.31, p < 0.05). However, scores for the Q250 condition were not always lower than for 500 pps. The highest-scoring subjects tended to maintain high scores for any rate condition.

In order to elucidate reasons for slightly poorer performance in the Q250 condition, a logistic regression analysis was performed on the aggregated voice gender categorizations across subjects for each stimulus: factors were the mean stimulus F0 (Kovačić and Balaban 2009), the voice gender, and the rate condition. The results are summarized in Fig. 8. Both F0 and speaker gender had a significant effect on gender classification (z = 10.78 and 4.82, p < 0.001). This indicates that both fundamental frequency and formant structure due to vocal tract length (VTL) had an effect on categorization; if this were not so, then the upper and lower regression lines in Fig. 8 would line up in the middle of the graph. There was also a significant interaction between gender and rate condition; contrast analyses revealed that for female voices, the 500 pps condition gave poorer performance compared to 1000 pps (z = 2.74, p < 0.01), likely due to F0 approaching the Nyquist rate (250 Hz). These rate effects were not observed for male voices (e.g., Q250 vs. 1000 pps, z = 1.40, p = 0.16, Fig. 8 lower curves).

Fig. 8.

Voice gender categorization versus F0 for three different rates: A logistic model was fitted to the counts of female and male responses across subjects, with F0, speaker gender, and rate as factors. Introducing an interaction between gender and rate slightly improved the model compared with no interaction. The curves indicate the modeled discrimination functions for female (right) and male gender (left), respectively: Female gender scores were significantly better at 1000 pps than for 500 pps. There were no other significant effects of rate. Several voices in the middle of the F0 range remained ambiguous or generally wrongly classified by subjects across the three rates (filled symbols). NB. Some points are horizontally separated for clarity

Overall, the effect of reduced rate conditions on gender classification performance was modest and could have been due to the limited exposure subjects received. We note that two of the stimuli had ambiguous gender (Fig. 8, solid shapes), as reported by Kovačić and Balaban (2009); removing these stimuli from the regression analysis did not alter the general pattern of results.

Results of Experiments 1 and 2 Compared

Seven CI subjects completed pitch ranking in both Experiments 1 and 2, and ten subjects voice gender categorization task. There were no significant differences in scores between ACE and STEP 1000 pps for any of the tasks (low F0 range: mean 1.0 vs 1.1, t[6] = 0.21, p = 0.84; high F0 range: 1.1 vs 1.1, t[6] = 0.05, p = 0.96; gender: 36.2 vs 37.0, t[9] = 1.24, p = 0.25). The loudness roving procedure used in Experiment 2 did not seem to reduce overall performance on the pitch-ranking task for STEP using 1000 pps compared with ACE using a similar rate and no loudness roving. Therefore, it appears that subjects were genuinely using pitch cues to make judgments of pitch rank rather than loudness cues. The integrator constant α, set to 0.5 for STEP in Experiment 2 produced similar performance when compared with those for ACE in Experiment 1. Choosing appropriate α reflects a trade-off between adequate envelope modulation bandwidth for optimal pitch perception and sound quality; as mentioned in the Methods section, the first author reported α = 0.5 to present a good compromise.

DISCUSSION

In this study, we compared voice pitch between the ACE and the STEP and investigated the effects of reduced carrier rate using STEP on voice pitch. Experiment 1 established equivalent performance between the clinical ACE coder and the novel STEP strategy at low fundamental frequencies. For high fundamental frequencies, performance for ACE was slightly better, but this result was largely driven by one participant (CI-18). The STEP was designed to control the spectral and temporal envelopes of the amplitude-modulated pulse trains used to convey information to the auditory nerve in CI users. Use of a single filter bank, as in ACE, introduces constraints on the temporal envelope characteristics, such as limiting the modulation bandwidth but can also introduce phase differences in the modulation produced by a given voice F0 across channels: this may reduce the salience of the modulation pitch cue for CI users. Conversely, if modulations corresponding to voice F0 are to be faithfully coded, for example, up to 200 Hz or above, then this imposes limitations on the spectral resolution that can be obtained with a single bank of filters. STEP overcomes these limitations while retaining ecological signal-driven information about the sound source rather than imposing synthetic modulation waveforms on the output signal, as used in other approaches such as F0mod (Milczynski et al. 2009; Francart et al. 2015) and eTone/OPAL (Vandali and van Hoesel 2012; Vandali et al. 2017). In this way, STEP may better preserve differences which occur between voices, as well as other differences which may result from variations in modulation waveforms.

Loudness Rating

We employed loudness scaling to eliminate potential loudness differences between coders and rates on perception. The highest attenuation setting used by subjects was set to − 15 dB in each experiment. Using Fig. 4a and the attenuation settings used to generate the histogram for ACE and STEP, even − 15 dB attenuation could have resulted in substantial saturation of the loudness growth function, with output stimulation being hard clipped and thus losing modulation depth. Potentially, a better method may be to customize MAP levels to compensate for loudness differences between coders, caused either by filter bank configuration or by coder rate.

It should be noted that the duration of pulse trains used to set MAP levels are of the order of one second, with constant amplitude, whereas coded sound is more dynamic with mostly relatively short transient peaks. Also, fitting based on the ACE strategy can reduce psychophysical MAP levels due to loudness summation. As we see in Fig. 4a and c, the dynamics of the output stimulation amplitudes can be altered by the filter bank or coder type, and therefore a different coder (or different parameters) may allow for higher tolerable MAP levels. Compensation for these loudness effects has also been required for other experimental sound coders looking to enhance F0 cues (Milczynski et al. 2009; Vandali and van Hoesel 2012). We saw a range of attenuator settings to achieve a loudness rating of 4.5 and, in some cases, MAP levels may have been artificially low in order for subjects to use their clinical ACE strategies in everyday settings (for example, where there is a low tolerance for background noise). Increasing MAP levels, rather than changing attenuator settings, could improve performance by reducing saturation of the amplitude mapping I/O function, as well as placing input signals in a less compressed region of the amplitude mapping function and therefore made modulations more salient.

In Experiment 2, we chose not to reset MAP levels according to the rate to limit the number of experimental variables. In general, clinical MAP levels decrease with increasing stimulation rates (Skinner et al. 1995; Zhou et al. 2012). We note that the results of the loudness ratings (e.g., Fig. 6a and b) indicate that there was no significant difference in loudness between Q250 and 500 pps conditions, which, to some extent, vindicates our approach. It is a larger question whether, for example, CI users could be given control over stimulation rate, and whether MAP levels should be co-varied or whether signal intensity should be varied to normalize loudness perception. From a signal processing perspective, dynamically varying such parameters as input attenuation or loudness growth functions would be simpler than varying MAP levels and using a single set of MAP levels also more straightforward for the clinician. Current Nucleus sound processors allow user modification of the loudness growth function via the “volume” control by reducing the varying maximum current level within the MAP level range.

Vowel Pitch Ranking and Voice Gender Categorization

Pitch ranking performance for the low F0 range was not generally affected by sound coder condition in either Experiment 1 or Experiment 2. Clearly, the lowest rate STEP Q250 condition supported voice pitch perception as well as the higher rate conditions (including those in Experiment 1). However, the performance was more variable for the higher F0 range, 164 Hz to 212 Hz, across sound coder rate conditions. Similarly, Fraser and McKay (2012) reported that modulation detection for single electrode stimuli was weak or absent for F0 > 150 Hz in a small group of CI users, and thus F0 discrimination would likely be impaired in this range. Milczynski et al. (2009) also saw no pitch-ranking ability in CI subjects using enhanced F0 coding for F0 > 200 Hz, and it is not clear from their testing in the 165 to 208 Hz range whether salient modulation cues were actually only available for the lowest, F0 = 165 Hz, stimulus since they only used a low-frequency reference F0 (we used reference F0s at each end of each range). In situations with F0 above 212 Hz, such as children’s voices or even adults expressing specific emotions, it is likely that the pitch is more dominated by place cues, with stimulation output patterns being shifted to more basal electrodes. We chose to focus on temporal cues to pitch and did not test this high F0 range, but the improved spectral resolution provided by STEP may enhance these place cues.

In Experiment 1, voice gender categorization was well above chance for all but one subject, in contrast to vowel pitch ranking, where only three subjects could consistently rank items. This finding may imply that performance on the pitch-ranking task does not indicate functional pitch perception required for gender categorization. However, the range of F0 used in pitch ranking represents approximately 2/3 of the entire male to female F0 range, with the low F0 range covering the overlap range for male and female F0. Given the overlap range of F0 for male and female talkers, subjects appeared to be using VTL cues to categorize voice gender. Gender test stimuli were sentence fragments of 2-s duration that may have allowed them to identify VTL characteristics (Meister et al. 2016) or, alternatively, to estimate the speakers’ range of F0, such as to identify the gender more easily. In the transition range of male to female F0, even a limited capacity to detect modulation at the lower F0 frequencies may in itself be a cue to gender, such that male voices produce a rough sound and female voices a smoother sound. This information may allow functional voice gender categorization despite overall poor temporal F0 perception.

Pitch ranking was not affected by coder type or rate for the low-frequency range, and a difference was only seen between 1000 and 500 pps for the high F0 range. We note that very few subjects who had systematic pitch ranking abilities were affected by the change to 500 pps rate. The effect of 500 pps was also only seen for female gender items, with higher F0s, which agrees with the pitch-ranking data. Again, as in Experiment 1, VTL cues may have assisted correct gender discrimination in the presence of degraded temporal F0 cues. Coding strategies designed to enhance temporal F0 perception (i.e. F0mod, eTone/OPAL) have generally employed higher rate carriers (> ≈ 1000 pps/channel) so as to reproduce the F0 cycle fully (e.g., Fig. 3a), as well as synchronizing temporal cues to F0 across channels. The current results indicate that these higher carrier rates may not be necessary for a large proportion of CI users, with the benefit that with reduced rates, battery life may be extended, and sound processors may be made smaller.

There may also be performance benefits from using lower carrier rates for temporal F0 perception, at least for some CI users: In Experiment 1, we saw one subject (CI8) with pitch reversals in the high F0 range using STEP, but good performance using STEP in the low F0 range. As the carrier rate of 900 pps should suffice to sample the F0 periods (similar to Fig. 3a, 1000 pps) up to the highest F0, we might hypothesize that the auditory nerve responses were not synchronized to carrier pulses due to refractory effects (Hughes et al. 2014), such that the neural carrier rate may have been half or some other fraction of the physical carrier rate. This would then produce aliasing in the peripheral neural domain and would, therefore, have an effect on the representation of the modulation signal similar to that obtained with a low fixed rate carrier (e.g., Fig. 3a, 250 pps). Chatterjee and Oberzut (2011), using amplitude modulation of pulse trains, found that the most sensitive CI users can detect subsampled patterns in the CI output leading them to perceive wrong pitch. Pfingst et al. (2011) established that peripheral cochlear health affects pulse train following by neural populations. The subject with good low F0 pitch ranking may have had appropriate central integration of modulation pitch cues, but peripheral integration limitations produced aliasing in the high F0 range (Hughes et al. 2014). The second author has also seen this pitch reversal pattern in a large group of subjects using clinical ACE processors consistent with findings in other studies (Kong and Carlyon 2010; Lamping et al. 2020).

In Experiment 2, we aimed to characterize the potential deficit in pitch ranking (using temporal F0 cues) and voice gender categorization, and as such, we chose subjects who generally performed best in Experiment 1. However, it may be that using a low fixed rate carrier of 500 pps or even lower rate F0 synchronized carrier like Q250 may have improved performance for poorly performing subjects, by overcoming peripheral bottlenecks in carrier rate. Thus, these conditions should be tested in the future, along with a synchronized condition running nominally 500 pps. Todd et al. 2019 recently described experiments to determine the use of temporal fine structure (i.e., carrier pulse timings) versus envelope timing cues in binaural unmasking in bilateral CI users. Their results indicated that temporal fine structure cues are dominant for low carrier rates, whereas envelope timing cues dominate for high carrier rates. These effects should be further evaluated for voice pitch perception, and STEP should be evaluated for bilateral and bimodal CI benefits.

Arora et al. (2011) investigated the effects of varying carrier rates on modulation detection at 50 Hz and 100 Hz, where loudness was balanced between programs. The modulation frequencies they tested represent the upper range for speech envelope modulation and the lower range of speech periodicity, or voice pitch (fundamental frequency F0, 50–500 Hz). Modulation detection was generally better for the 500 pps carrier compared with a higher rate (900 pps) carrier when modulation depth was expressed in terms of current amplitude. However, because increasing carrier rate generally increases loudness, sound processor MAP levels, comprising of subjects’ clinical threshold (T) and comfort (C) levels, also vary as a function of rate (Skinner et al. 1995; Zhou et al. 2012). When Arora et al. (2011) expressed the results in terms of acoustic input sound level, only the lowest rate, 275 pps-carrier, produced significantly higher modulation thresholds. Speech recognition in noise was also significantly worse for the 275 pps carrier compared with 500 pps and 900 pps carriers, but this was not so for word recognition in quiet. Overall, these results suggested that a moderately low carrier rate may be sufficient for representing speech signal modulations up to 100 Hz.

Finally, the small differences in performance between rate conditions found in the current experiments may be partially explained by subjects’ having only acute exposure to the low-rate conditions. Thus, future work in this area should include take-home experience with the novel coder using low-rate conditions. We would also envisage that the effects of variations in frequency to electrode allocation on voice pitch perception can be further explored using such a novel coder.

Funding

This study was supported by the Investigator-Initiated Research (IIR) Grants #2098 and #853 from Cochlear© awarded to DK. Besides, DK was also partially supported by the EU structural grant for Croatia (code: HR.3.2.01–0320) and Split-Dalmatia County.

DECLARATIONS

Ethics Approval and Consent to Participate

The study was approved by the ethical committee of the SUVAG Polyclinic (# 510–08/17–20). Either CI subject or one of their parents gave written informed consent, depending on the subjects’ age.