Scientific methods / version 1.0.0
Cochlear Implant Laboratory
Methods & Validation
← Open the laboratory · Anatomical models · Implant references · Walkthrough · Neural model · Numerical evidence · References
The laboratory calculates a sound-processing chain, a sequential electrical pulse schedule, finite-contact extracellular fields, and the resulting membrane dynamics of multicompartment auditory-nerve fibers. Its electrodogram, membrane traces, spike raster, and recruitment comparisons come from those calculations.
The implemented nerve is an independent, deterministic implementation of the published 65-compartment Rattay human auditory-nerve model. Its population is arranged in an idealized, unrolled homogeneous conductor. The anatomical viewer combines measured human chamber surfaces, reconstructed microanatomy, and a display mapping of those calculated results. The release does not predict a person’s hearing, speech recognition, perceived pitch, comfortable current, surgical outcome, or long-term adaptation.
Anatomical models: measured surfaces and reconstructed detail
The whole-cochlea and cutaway views use the OpenEar ZETA human specimen, released by Sieber and colleagues under CC BY 4.0. OpenEar registers cone-beam CT with microscopic serial sections; it is not a generic spiral mesh. The scala tympani, scala vestibuli / vestibular labyrinth, and cochleovestibular nerve surfaces share their original specimen registration. Sources: methods paper and dataset and attribution.
For interactive display, the source meshes are simplified and transformed by one rigid, right-handed rotation and translation, in millimetres. The display removes distant vestibular structures from the scala-vestibuli / labyrinth mesh; the cutaway additionally removes facing triangles to expose the interior. These operations preserve the retained surfaces’ measured coordinates. The cutaway opening is an inspection window, not a segmented anatomical boundary. A failed atlas download is identified in the viewer and falls back to reconstructed lumen surfaces.
| Structure | Representation | Interpretation |
|---|---|---|
| Scala tympani, scala vestibuli, nerve trunk | Measured OpenEar ZETA surfaces, simplified for display. | One human specimen; not an average or a patient-specific model. |
| Scala media, basilar and Reissner’s membranes, modiolus, spiral lamina | Reconstructed surfaces registered around the measured scala-tympani curve. | Their detailed boundaries are not independently segmented from this asset. |
| Organ of Corti | A separate cellular-scale reconstruction with one inner-hair-cell row, three outer-hair-cell rows, supporting cells, the tunnel of Corti, reticular lamina, and tectorial membrane. | Selected structures and representative dimensions illustrate organization; this is not a complete histological volume. |
| Stereocilia and afferent endings | Stepped bundles, radial type-I endings at inner hair cells, and sparse type-II outer-spiral routes. | Hair-cell transduction, electromotility, synaptic release, and type-II membrane dynamics are not calculated. |
| Type-I fibers and spiral-ganglion somata | Curved display paths with separate bare somata, process myelin, and nodes. | A representative subset includes the selected modeled fiber. Its compartment colors can replay the existing cable result; the drawn route does not replace the solver geometry. |
Reconstruction choices draw on human organ-of-Corti microscopy, basilar-membrane morphology, and auditory-nerve ultrastructure. The cellular patch samples ten hair cells along each displayed row and 27 stereocilia per cell. That sampling illustrates bundle organization rather than claiming a complete stereocilium count.
The measured scala-tympani medial curve has a reconstructed path length of 22.247 mm. It is a lumen path, not a measurement of basilar-membrane length. The numerical model’s fixed 35 mm place coordinate maps by normalized arc length onto this curve for display only. Anatomical distances in the viewer must not be substituted for calculated contact-to-neuron distances. The cellular view uses its own micrometre-scale bar; whole-cochlea neurites and somata are enlarged to remain visible.
Asset checks compared the simplified meshes with the deposited originals. All three complete simplified surfaces remain watertight. The scala-tympani volume change is -0.0372%; the largest surface distance among the bidirectional samples is 5.32 µm. All 1,981 tested centreline samples lie inside the original chamber. These are finite-sampling checks against a source segmentation, not biological-accuracy measurements or a global surface-error bound.
The Auditory neuron view isolates the selected model’s ganglion soma and adjoining processes. It retains the cable model’s compartment lengths and diameters, including the segmented 30 µm soma, and crops long process continuations to keep this junction inspectable. Process myelin appearance and intracellular / glial context are reconstructed; those static context surfaces do not acquire membrane voltages. The exposed membrane surface can replay the exact stored compartment values.
Rotate by dragging, zoom with the wheel or the camera buttons, and use the layer controls to expose structures. Contact and nerve selection use surface picking. A software-rendered fallback keeps the same three-dimensional meshes and camera controls available when WebGL cannot initialize.
Implant construction: five sourced reference models
The device-inspection view represents published physical dimensions and contact topology for five electrode arrays. Each has a tapered carrier, individual metallic stimulating pads, relevant insertion markers, and a reconstructed unimplanted pose. These are independent dimension-based reconstructions, not manufacturer CAD. Intermediate taper, exact pad outlines, and undocumented free-rest curvature remain approximations; the viewer retains the source notes for each device.
Channel positions, physical pads, electronic current sources, and current-steered virtual sites are different quantities. In particular, FLEX28 has 12 longitudinal channel positions and 19 physical pads: five apical single pads and seven paired sites. The Mid-Scala radiopaque marker is inactive and is not counted as a stimulating contact. Inspection-site order runs from apex to base; it is independent of both manufacturer numbering and the experiment’s basal-first source indices.
Bench inspection retains physical carrier dimensions. Inserted inspection places the array by physical arc distance along the measured reference curve and applies an conservative clearance estimate sampled from the original scala-tympani surface. If the requested placement exceeds the path or clearance, the model leaves the remaining length outside the entry; it does not shrink the device to force insertion. These geometric checks do not model insertion forces, tissue deformation, trauma, or clinical suitability. A lateral-wall, mid-scala, or perimodiolar family describes the device’s design; the generic path placement is not a validated reconstruction of that device’s surgical position.
The “Use channels in the experiment” action copies the site count, enables all source channels, and selects a feasible generic eight-maxima processor. The reference span, reference position, source-disk model, and unrolled conductor remain the experiment’s settings. The vendor inspection mesh has no electrical replay, manufacturer-specific field solution, or proprietary processor assigned to it.
1. A first experiment
- Choose “Follow a tone.” Inspect the acoustic waveform and the settings for the 1 kHz signal, contact array, and source current.
- Run the experiment. The sound processor generates pulses, the field solver evaluates their extracellular effect, and the active cable model calculates each surviving fiber’s response.
- Select a contact and a fiber. The electrodogram shows delivered source pulses. The membrane panel shows voltage in the selected fiber’s peripheral node, soma, and distal axon. The raster counts detected distal spikes.
- Scrub the replay. Replay displays stored solver frames; it does not change simulation time, pulse rate, or the numerical integration.
- Pin a reference and change one condition. Distance, current, return configuration, and source-channel availability act on different parts of the calculation. Compare the settings as well as the output.
- Use the diagnostic tools. The single-contact scan measures isolated-pulse recruitment overlap. The threshold search brackets a deterministic model response for one contact and one fiber. Export settings, pulses, spikes, or the selected membrane trace to inspect the result outside the interface.
Changing controls prepares a new experiment. A completed result retains the settings that generated it; run the model again before interpreting the new configuration. An uploaded audio clip supplies the selected neural window. Synthetic sources are explicitly labeled, including the harmonic, formant-shaped “Synthetic vowel” example.
2. Seven comparison experiments
| Experiment | What to do | How to interpret it |
|---|---|---|
| Follow a tone | Start with the 1 kHz tone. Follow its filter-bank envelope into the delivered pulse schedule, then select a fiber and inspect peripheral-node, soma, and distal-axon voltages. | A high envelope, a large extracellular potential, and a propagated action potential are different observations. The model can show any of them without all three being equivalent. |
| Frequency sweep | Use the logarithmic sweep with four maxima. Scrub time to follow the changing input frequency through selected contacts and the neural raster. | This preset compresses a sweep into a short numerical window. Its two-second listening counterpart traverses the same frequency range more slowly. |
| Two frequencies | Compare the 800 Hz and 2,800 Hz inputs, their selected source contacts, and the recruited fiber locations. Use the isolated-contact scan to inspect overlap. | Two spectral peaks do not guarantee two independent recruited populations. The overlap scan uses separate single pulses, while the sound experiment is a pulse train. |
| Closer array | Pin the tone baseline, then reduce the model contact-to-fiber separation from 0.5 to 0.25 mm at the same source amplitude. | Distance changes the actual lead field. Equal current is the comparison condition; perceived loudness and clinical benefit are not matched. |
| More contacts | Place 24 contacts across the same 16 mm reference span, with eight maxima. For a controlled comparison, also set the 12-contact reference to eight-maxima n-of-m. | The preset changes both contact count and the baseline coding strategy. Match the strategy first when asking specifically about contact count; more contacts do not automatically mean more independent information. |
| Local nerve loss | Remove modeled fibers within the specified reference region while retaining the same acoustic input and source schedule. | This is a seeded survival mask on otherwise identical model fibers. It does not simulate degeneration, altered excitability of surviving fibers, or an individual recipient’s neural survival. |
| Focused return | Compare monopolar and partial-tripolar returns using the same pulse amplitudes, widths, and contact geometry. Inspect absolute recruitment alongside field shape. | A narrower normalized field can accompany reduced recruitment. The comparison does not establish equal loudness, equal energy, or improved speech understanding. |
The additional Synthetic vowel preset is a generated harmonic signal with broad formant-like peaks. It is useful for inspecting maxima selection and listening to the vocoder. It contains no recorded speech and does not constitute a speech-understanding test.
3. Geometry and frequency-place assignment
The quantitative coordinates are in millimeters: x is position from the base along a fixed 35 mm reference path; y follows the straight compartment chain of a nerve fiber; and z is contact-to-fiber separation. Circular contacts lie in the x–y plane, and the population’s fiber centerlines lie at z = 0. All surviving fibers use the same membrane and compartment parameters.
Contact 1 is the most basal, highest-frequency source channel. The final contact is the most apical, lowest-frequency channel. Contacts are equally spaced over the chosen reference span. The apical-position control translates that span along the reference coordinate. These distances are not a conversion from a manufacturer’s physical electrode-array length or a CT-measured insertion depth.
The displayed place-frequency relation uses the human coefficients of Greenwood’s function with the declared 35 mm reference length:
f(x) = 165.4 × [10^(2.1 × (1 − x / 35)) − 0.88] Hz mismatch = 12 × log₂(place frequency / assigned band center) semitones
The mismatch is a geometric reference calculation. It is not a direct perceived-pitch estimate or a patient-specific spiral-ganglion map. Anatomical mapping between the organ of Corti and spiral ganglion requires additional geometry, as demonstrated by Stakhovskaya and colleagues. The anatomical viewer maps this place coordinate onto the measured specimen for display. Its bone, fluid spaces, and curved neurites do not enter the electrical calculation.
4. Sound processing and pulse scheduling
The transparent processor implements generic continuous interleaved sampling (CIS) and generic n-of-m selection. It does not claim to reproduce a manufacturer’s complete ACE, HiRes, or other proprietary processing chain. Audio levels are normalized digital amplitude and dBFS; they are not dB SPL unless an external calibration is supplied.
Analysis bands have logarithmic boundaries. Each band uses two cascaded second-order bandpass sections with prewarped edge frequencies; the complete fourth-order response is −3 dB at each declared band boundary. Full-wave rectification is followed by a fourth-order Butterworth low-pass filter. Multiplication by π/2 gives a peak-equivalent envelope for a steady sinusoid. The onset transient is part of the causal calculation.
eᵢ(t) = (π / 2) × LPF(|BPFᵢ(audio)|) uᵢ = clamp((20 log₁₀(eᵢ) − input minimum) / input range, 0, 1)^γ Iᵢ = minimum current + uᵢ × (maximum current − minimum current)
Audio at or below the input minimum produces no pulse. CIS schedules each eligible source channel in a frame. N-of-m ranks the pre-compression envelopes of enabled sources, selects up to the requested number of maxima, and schedules them in basal-to-apical order. Disabled sources cannot occupy a maximum. Equal minimum and maximum currents define a fixed-current test; zero for both defines a true no-stimulation control.
The selection-frame rate is distinct from total pulse rate and from a particular contact’s achieved rate. Under n-of-m, contact selection varies with the signal. The total possible pulse rate is selected contacts per frame multiplied by frame rate; the actual per-contact rates are counted from delivered events. Gray and colleagues explicitly describe this distinction in measurements of commercial processors.
Timing and charge
Every scheduled event is a symmetric biphasic rectangular pulse with a zero-current interphase gap. A cathodic-first pulse starts negative at the named source and has an equal positive return phase. The opposite leading polarity reverses both signs. Pulse onset timestamps are rounded to a 1 ns grid, independently of the audio-sample clock. This timestamp precision is not a claim of 1 ns accuracy in the neural integration.
Qphase [nC] = amplitude [µA] × phase width [µs] / 1000 Qnet = Qfirst + Qsecond = 0 occupancy = K × frame rate × (2 × phase width + gap + guard) × 10⁻⁶
Here K is the enabled-source count for CIS, or the smaller of maxima and enabled-source count for n-of-m. Occupancy must not exceed one. A 12-source CIS schedule at 800 Hz with 25 µs phases, an 8 µs gap, and a 2 µs guard occupies 57.6% of the available time; at 1,600 Hz it requires 115.2% and is rejected. A 200 µA, 25 µs phase carries 5 nC.
Charge balance describes the prescribed waveform. Electrode polarization, impedance, compliance limits, electrochemical reactions, and tissue injury are not modeled. The simulator therefore reports current and charge without converting them into clinical comfort or safety claims. The joint acoustic-input and electrical-parameter checks follow the kind of explicit evaluation advocated by Shekar and Hansen.
5. Finite-contact fields and returns
The field is the full-space Green-function solution for an unbounded, homogeneous, isotropic conductor. Each stimulating contact is a circular disk with prescribed uniform source density. It is not an equipotential metal-contact boundary solution. Conductivity is constant throughout the calculation; there are no scala, bone, modiolus, or insulating tissue boundaries.
φ(r) = I / (4πσA) × ∫disk 1 / |r − r′| dA′ φon-axis(z) = I / [2πσ(√(z² + a²) + |z|)]
The normal-axis expression is evaluated analytically. Off-axis potentials use Gauss–Legendre radial quadrature and angular quadrature over the disk. With current in µA and distances in mm, the numerical result is mV. The far field approaches I/(4πσr). An insulating half-space would have a different factor and is not the model used here.
Linear superposition builds one field basis per possible source channel. Monopolar stimulation uses the source disk and an explicit remote point return. Bipolar stimulation returns current through the neighboring contact. Partial-tripolar stimulation splits the local return fraction across adjacent contacts, with the remaining fraction returning remotely. At array ends, the unavailable adjacent share also returns remotely. The coefficients sum to zero for every configuration.
Source-channel disabling changes eligibility to emit a scheduled source pulse. A disabled source channel may still serve as a local return for a neighboring source. It is not a model of an electrically disconnected electrode. The remote return is located at [17.5, −20, 20] mm in the unrolled coordinates.
The field view reports potential gain in mV/µA for a selected source/return arrangement. The active nerve is driven by the spatial variation of this extracellular potential along its compartments. A broad potential contour is not itself a recruitment contour.
6. Active auditory-nerve model
The independent implementation follows the 65-compartment Rattay et al. model, cross-checked against Bachmaier et al.’s comparison and author implementation. It includes the peripheral terminal and myelinated peripheral process, presomatic segments, a ten-segment soma, postsomatic segment, and myelinated central axon with active nodes. Compartments use truncated-cone surface areas and integrated axial half-resistances.
Voltage is intracellular minus extracellular potential. Active compartments contain sodium, potassium, and leak currents with voltage-dependent Hodgkin–Huxley-type gates. Myelinated compartments and the soma retain the published membrane scaling. All fibers begin at −65 mV with equilibrium gates; there is no added spontaneous drive, stochastic channel noise, synaptic model, or long-term adaptation mechanism.
Cᵢ dVᵢ/dt = −Iion,ᵢ
+ Σⱼ gᵢⱼ[(Vⱼ − Vᵢ) + (φₑ,ⱼ − φₑ,ᵢ)]
Iion,ᵢ = gNa,ᵢ mᵢ³hᵢ(Vᵢ − ENa)
+ gK,ᵢ nᵢ⁴(Vᵢ − EK) + gleak,ᵢ(Vᵢ − Eleak)
+ gmyelin,ᵢ(Vᵢ − Vrest)
dx/dt = αx(V)(1 − x) − βx(V)x The same axial conductances act on intracellular and extracellular voltage differences, including sealed endpoints. A spatially uniform extracellular potential therefore produces no cable-driving current. Refractory effects arise from the active channel dynamics; no arbitrary “one spike per pulse” rule forces responses.
One implementation difference is retained explicitly: this laboratory uses the integrated truncated-cone axial conductances for both membrane coupling and extracellular drive. The author reference stimulus generator uses midpoint-diameter resistance approximations for its extracellular drive. Small waveform differences are therefore expected even when the underlying published membrane parameters agree.
Numerical solution and recorded events
The solver updates gates with an exponential voltage-held step and solves the stiff cable with a backward-Euler tridiagonal solve. Steps are shortened to land on pulse edges and recording times. A 0.025 mV gate table accelerates full steps; exact rate calculations are used outside the table and for shortened steps. The default integration step is 1 µs. Integration timestep, trace interval, and replay interval are separate settings.
The laboratory detects an upward 0 mV crossing at a declared distal axonal node, rearming below −40 mV. The numerical reference comparison also preserves the author comparison’s −5 mV criterion when required. The nerve window extends 3 ms past the acoustic input window to retain delayed responses. A displayed spike is a detected event at that modeled compartment. It is not an auditory brainstem response or an electrically evoked compound action potential; those require an additional recording forward model.
The survival control removes a seeded subset of model fibers in a specified reference-place interval. The surviving fibers’ geometry and membrane parameters remain unchanged. No adaptation, degeneration time course, cortical learning, or recipient-specific calibration is inferred from this intervention.
7. Recruitment, overlap, and threshold tools
Spatial recruitment counts a fiber once if it has at least one distal spike in the run. The isolated-contact scan separately delivers one equal-amplitude biphasic pulse per enabled source, using the current waveform and return settings, and records each source’s recruited set.
cosine overlap(i,j) = |Aᵢ ∩ Aⱼ| / √(|Aᵢ| × |Aⱼ|)
A value of zero denotes disjoint recruited sets; one denotes identical nonempty binary recruitment profiles. A pair is undefined when either source recruits no fibers, and its cell remains blank. This is cosine overlap, not Jaccard overlap. It describes this deterministic, isolated-pulse experiment; it does not measure independent perceptual channels or predict speech scores. Goehring et al. and Bierer and Litvak motivate examining channel interaction without providing a universal conversion from this score to hearing performance.
The threshold tool first samples increasing currents in 25 µA steps up to 2,000 µA, then bisects the first sampled nonfiring-to-firing bracket ten times. Its response criterion is a distal 0 mV upcrossing within a 4 ms single-pulse experiment. The returned interval is a local numerical bracket. Sampling can miss a narrow nonmonotonic response region; failure to find a response below the ceiling is reported as no bracket found, not as proof of permanent inexcitable tissue. This model threshold is not a person’s hearing threshold or comfortable level.
8. Acoustic previews
The sine and noise vocoders resynthesize selected acoustic envelopes using frequency-specific carriers. They can illustrate analysis-band limitations, maxima selection, and frequency shifts. Modulation uses the acoustic envelopes gated by delivered-source selection; it does not interpret electrical current as acoustic sound pressure. These acoustic demonstrations do not convert the computed nerve spikes into a validated percept.
Synthetic listening previews use a two-second version of the source, while the neural experiment uses its explicitly selected shorter window. Their durations and, for a sweep or evolving vowel, their time courses differ. The neural raster is not a two-second solve. Uploaded audio previews use the available clip according to the displayed controls.
Vocoder gain normalization is explicit: the default targets an RMS of 0.12 in digital amplitude and limits the peak to 0.9 before the playback gain. Noise is seeded for repeatability. A large upward carrier shift can move carriers beyond the available audio bandwidth; those carriers are omitted. The carrier-shift control changes this acoustic synthesis, not electrode position, the finite-contact field, or the neural result.
For amplitude-modulated tones, the source envelope is proportional to (1 + m sin(2πfmt)) / (1 + m), where m is modulation depth from zero to one. Zero depth gives the unmodulated carrier. Generated waveforms include short onset/offset ramps and are normalized to their specified peak digital level.
Listen to the preview as an illustration of coding constraints. Direct comparison studies in single-sided-deaf listeners show that acoustic approximations vary with the individual and chosen parameters; a generic vocoder should not be labeled as what every implant user hears. See Dorman et al. and Valid Acoustic Models of Cochlear Implants.
9. Numerical evidence and reproducibility
The verification separates analytical field checks, processing and timing checks, end-to-end invariants, and comparisons against an independent implementation of the published nerve model. Numerical agreement verifies the specified computation. It does not establish biological or patient validation.
| Check | Executed result | Meaning |
|---|---|---|
| Off-axis disk potential | Eight fixtures compared with an independent elliptic-integral and adaptive radial calculation. Worst relative error: 0.005812% at the default 8 × 24 quadrature; 2.23 × 10⁻¹⁰ at 16 × 64. | Tests the declared uniform-source disk field, including units and off-axis integration. |
| Processor and audio | Analytical filter-edge responses, tone mapping, silence, zero-current control, source masks, maxima, timing feasibility, phase charge, deterministic vocoding, and AM-depth behavior are checked. | Tests actual numerical and scheduling outputs, rather than a display-only response. |
| Published thresholds | All six selected published threshold fixtures bracketed within ±0.5% at the author implementation’s 5 µs timestep. | Reproduces these reference configurations and criterion; it does not establish an individual clinical threshold. |
| Independent author-code comparison | At 1 µs, three single-/multiple-pulse cases had five-trace RMSEs of 0.01571, 0.01053, 0.01132 mV; the largest absolute trace discrepancy was 0.17437 mV. | Compares the independent JavaScript implementation with the original author Python model running in Brian2 2.10.1. |
| Time refinement | Against a 0.25 µs reference in the sampled three-pulse experiment, trace RMSE was 1.6836 mV at 2 µs, 0.7159 mV at 1 µs, and 0.2372 mV at 0.5 µs. The largest pointwise voltage difference at 2 µs was 17.18 mV, with a maximum spike-time difference of 11.75 µs. | Retains the observed discretization error. Steep spike edges amplify small latency differences in pointwise voltage comparisons; event-timestamp precision does not remove integration error. |
| Gate interpolation | The tested 0.025 mV table differed from exact gate calculations by approximately 0.0009732 mV in the trace and 0.0005300 µs in spike timing. | Checks the acceleration approximation in the sampled configuration. |
| Integration controls | Independent tests exercise rest, uniform extracellular potential, finite-contact-driven propagation, source disabling, distance and survival changes, overlap calculations, and saved-setting reproducibility. | Checks that the processing, geometry, field, and cable stages remain connected. |
These figures describe the retained verification fixtures, not error bounds for every user-selected experiment. The browser’s deterministic output depends on the full configuration, audio samples, seed, solver resolution, and model version. Sensitive comparisons should be repeated at a finer timestep; changes in fiber count or reference geometry are separate changes to the modeled experiment.
The source repository retains the validation record, quantitative neural comparison, author-run reference traces and source identifiers, and automated tests. The original author model is used for comparison; its source code is not redistributed as the laboratory’s implementation.
node --test tests/cochlear-*.mjs node scripts/cochlear-neural-validation.mjs python scripts/cochlear-neural-reference.py --author-source PATH
10. Primary sources
Sieber et al. (2019). The OpenEar library of 3D anatomical models of the human temporal bone. Registered cone-beam CT and serial micro-slicing; source of the measured scala surfaces.
Sieber et al. — OpenEar dataset, ZETA specimen. OpenEar library, Zenodo record 1473724. CC BY 4.0 source meshes; retained in their shared specimen coordinate frame before a rigid display transform.
Wright (1983). Scanning electron microscopy of the human organ of Corti. Human hair-cell rows, stereocilia, and sensory-epithelium organization.
Liu et al. (2015). Macromolecular organization and fine structure of the human basilar membrane. Human membrane morphology and increasing width toward the apex.
Liu et al. (2021). Spike Generators and Cell Signaling in the Human Auditory Nerve: An Ultrastructural, Super-Resolution, and Gene Hybridization Study. Human soma, initial-segment, satellite-glial, and process-myelin anatomy.
Rattay, Lutter & Felix (2001). A model of the electrically excited human cochlear neuron. I. Contribution of neural substructures to the generation and propagation of spikes. Source of the human auditory-nerve compartment model.
Bachmaier et al. (2019). Comparison of Multi-Compartment Cable Models of Human Auditory Nerve Fibers. Parameter cross-check and independent published-model implementation used for numerical comparison.
Greenwood (1990). A cochlear frequency-position function for several species—29 years later. Basis of the declared frequency-place reference.
Stakhovskaya et al. (2007). Frequency Map for the Human Cochlear Spiral Ganglion: Implications for Cochlear Implants. Shows why organ-of-Corti place and spiral-ganglion geometry require distinct anatomical treatment.
Wilson et al. (1991). Better speech recognition with cochlear implants. Original continuous interleaved sampling experiment.
Gray et al. (2021). Transmission of Binaural Cues by Bilateral Cochlear Implants: Examining the Impacts of Bilaterally Independent Spectral Peak-Picking, Pulse Timing, and Compression. Empirical processor measurements and the distinction between selection-frame rate and total delivered pulse rate.
Shekar & Hansen (2021). An evaluation framework for research platforms to advance cochlear implant/hearing aid technology: A case study with CCi-MOBILE. Motivates explicit testing of acoustic inputs and realizable electrical pulse parameters.
Goehring et al. (2021). The effect of increased channel interaction on speech perception with cochlear implants. Motivates channel-interaction experiments; its perceptual findings do not calibrate this simulator’s overlap score.
Bierer & Litvak (2016). Reducing Channel Interaction Through Cochlear Implant Programming May Improve Speech Perception: Current Focusing and Channel Deactivation. Primary experiments on focusing and channel deactivation, with variable responses across participants and configurations.
Dorman et al. (2020). Approximations to the Voice of a Cochlear Implant: Explorations With Single-Sided Deaf Listeners. Direct comparisons between implant and acoustic sound quality.
Valid Acoustic Models of Cochlear Implants (2021). One Size Does Not Fit All. Demonstrates the importance of individual acoustic-model settings; a generic vocoder is not a universal perceptual replica.