PaperlayerView exact revision on arXiv

Source-linked AI summary

No speech articulator robustly favours retroauricular electrodes over canonical jaw sites: a volume-conductor model with orientation and electrode-count controls

Carl Vincent Kho

arXiv:2610.03723v1cs.HCeess.SP

TL;DR

The paper asks where electrodes should be placed to capture speech-articulator activity, a coupling not previously computed for anatomically resolved muscle conductors. It uses reciprocity on MIDA to compare electrode sites and articulators; the reported comparison favors the jaw for seven muscles and the ear for three, while the temporalis result depends on orientation treatment.

  • Problem

    The coupling between speech articulators and jaw- or ear-worn electrodes had not been computed, despite ear-worn arrays recording jaw and speech activity.

    Source: §1 · p. 2

  • Method

    The study uses reciprocity on MIDA to compare 22 electrode positions with ten segmented articulators while testing orientation and electrode-count controls.

    Source: §3.1 · p. 7

  • Results

    Seven of ten articulators couple more strongly to the jaw montage and three to the ear; temporalis’ derived fibre field gives an ear preference of 1.147 dB with an interval of [−1.453, +5.458].

    Source: §1 · p. 2, §4.1 · p. 11

  • Takeaways & Limitations

    A spatial filter informed by which muscle dominates at each contact differs from one that treats retroauricular electromyography as a single nuisance component.

    Source: §4.5 · p. 16

    The anatomically constrained orientation sweep is derived for temporalis and cannot be extended to the other nine muscles, which are swept uniformly over the hemisphere.

    Source: §4.7 · p. 18

Abstract

from arXiv · show

Objective. Ear-worn biopotential devices are being designed around a coupling that has not been computed: how strongly each speech articulator reaches electrodes on the jaw and around the ear. We compute it, and test the answer against the three things that could produce it spuriously - source orientation, electrode count, and the level of anatomical detail in the volume conductor. Approach. Articulator-to-electrode coupling is computed by reciprocity on MIDA, a head model with 116 labelled compartments at 500 microns, treating muscle as both the generator and its own conducting compartment. Twenty-two electrode positions - spanning the canonical jaw montage, a retroauricular cluster, and a cEEGrid C-path - are compared against ten individually segmented articulators. Source orientation is swept over the hemisphere, not assumed, and every comparison is repeated at matched electrode counts. Results. Five articulators - orbicularis oris, buccinator, mentalis, depressor anguli oris, and platysma - favour the jaw montage at every sampled orientation and at every electrode subsample. No muscle robustly favours the retroauricular montage. The strongest ear-leaning candidate, temporalis, reaches -2.57 dB under a uniform orientation sweep; but derived from the label volume, its fibre field gives -1.15 dB with an interval of [-1.45, +5.46], and only half of four-site retroauricular subsets favour the ear. Significance. For device design the result is one-sided, not a trade: a retroauricular montage loses lip and chin activity entirely and buys no muscle back reliably in exchange. What remains a design lever is placement - a four-site cluster chosen by anatomical target outperforms arbitrary placement around the ear by up to 1.07 dB.

1. Introduction

The paper addresses the uncomputed coupling between speech articulators and jaw- or ear-worn electrodes, asking where a sensor should be placed rather than how to model muscle sources. It applies reciprocity to anatomically resolved muscle compartments and examines how montage comparisons depend on their assumptions and controls.

1. Introduction

  • Devices are being designed around an uncomputed coupling between speech articulators and electrodes, making electrode placement the open design question.Ear-worn arrays already record jaw and speech activity, while the canonical montage places electrodes on the chin, under the jaw, and along the throat.Source: §1 · p. 2
  • No prior forward model treats facial and cervical muscle as both the generator and its own anatomically resolved conducting compartment.HArtMuT is a methodological precedent, but its muscle sources radiate through a homogeneous scalp compartment and were built for muscle-artifact removal from scalp EEG.Source: §1 · p. 2
  • The study uses reciprocity on MIDA to compare 22 jaw, retroauricular, and cEEGrid electrode positions with ten segmented articulators.It examines isotropic and anisotropic muscle conductivity and assembles uncertainty from measured terms.Source: §1 · p. 2
  • Seven articulators couple more strongly to the jaw montage and three to the ear, indicating that the montages see different muscles rather than forming a single ranking.The ear-leaning muscles named in this comparison are temporalis, sternocleidomastoid, and lateral pterygoid.Source: §1 · p. 2

2. Methods

The methods use MIDA anatomy and reciprocity to compare muscle-to-electrode coupling, with orientation, electrode placement, and model validation handled explicitly. The pipeline also documents uncertainty and reproducibility constraints, including which anatomical targets cannot be resolved separately.

2.1 Head model

  • Reciprocity on MIDA’s 116-label voxel model compares ten individually segmented articulators, while pooled tongue and suprahyoid muscles are absent from per-muscle results.The model is segmented at 500 µm; MIDA’s temporalis label also merges temporalis with temporoparietalis.Source: §2.1 · pp. 2–3

2.2 Conductivity assignment

  • Conductivities draw on SimNIBS defaults and the IT’IS database, while an air-conductivity sweep found departures no greater than 0.069 dB across 10^-6 to 10^-4 S/m.The chosen 10^-6 S/m air value lies within the reported insensitive range; at 10^-3 S/m, departures reach 0.467 dB.Source: §2.2 · pp. 3–4

2.3 Electrode placement

  • All 22 electrode sites are derived from labelled anatomy and projected to skin, using centroid targets when interior and minimum-distance targets otherwise.A throat site over sternocleidomastoid is withheld because truncation biases its centroid and prevents defensible automatic placement.Source: §2.3 · p. 4

2.4 Reciprocity

  • Reciprocity obtains a muscle source’s lead field by projecting the electrode-generated reciprocal field onto source orientation, requiring 22 electrode solves rather than roughly 10^5 source solves.Within each compartment, the reported field summary is the volume-weighted median magnitude.Source: §2.4 · pp. 4–5

2.5 Fibre orientation; 2.5.1 The anatomically-constrained sweep

  • Muscle orientation is bounded rather than assumed: PCA-derived anisotropic tensors apply to only two of ten muscles, and an anatomically constrained orientation sweep is available for temporalis alone.Other muscles are swept uniformly over the hemisphere; the temporalis sweep intersects that space with its anatomically permitted fibre directions.Source: §2.5 · p. 5, §2.5.1 · p. 5

2.6 Validation; 2.7 Error budget

  • Validation combines an analytic four-layer sphere, head-mesh reciprocity and physical invariants; the three-point, three-parameter convergence fit is consistency evidence, not corroboration.The error budget admits ratio-varying terms and separately records four directional but currently unquantifiable terms; uniform scale factors cancel in site ratios.Source: §2.6 · pp. 5–6, §2.7 · p. 6

2.8 Reproducibility and pre-registration

  • The preregistered anatomical prediction was retained despite not being confirmed, while exact-value reproduction requires the specific mesh that is not version-controlled.A clean checkout can rebuild the pipeline but not the exact mesh used for the reported values.Source: §2.8 · p. 6

2.9 Ethics

  • This was a computational study using a licensed anatomical model, with no human subjects, new imaging, recordings, or living-participant measurements.No IRB review was required or sought.Source: §2.9 · p. 7

3. Results

With source orientation and electrode count controlled, five labial and chin articulators robustly favour the jaw, while no muscle robustly favours the retroauricular montage. The remaining results show that anisotropy, tissue contrast, truncation, conductor detail, and electrode placement qualify magnitudes or design choices without overturning that central finding.

3.1 Which montage sees which muscle

  • Five articulators favour the jaw at every sampled orientation and electrode subsample; no fibre direction or four-site selection makes the ear competitive for them.The five are orbicularis oris, buccinator, mentalis, depressor anguli oris, and platysma.Source: §3.1 · p. 7
  • Temporalis does not robustly favour the ear: its derived-fibre median is −1.147 dB, but the matched-count interval [−1.453, +5.458] crosses zero.Only 50.5% of four-site subsets favour the ear; a uniform orientation sweep instead gives −2.571 dB with an interval excluding zero.Source: §3.1 · pp. 7–8

3.2 Anisotropy changes the field but not the comparison

  • Anisotropy changes site-to-site gaps by less than the 0.27 dB meshing floor, although it raises medial-pterygoid absolute lead field by roughly 5 dB.The tensor was applied to sternocleidomastoid and medial pterygoid; anisotropy affects jaw and ear sites alike, so its effect subtracts out of the reported ratios.Source: §3.2 · p. 9, Figure 4 · p. 9

3.3 The tissue-conductivity contrast is a small term with a muscle-dependent sign

  • Tissue contrast has a muscle-dependent sign: it builds 2.86 dB of the mentalis gap but suppresses 1.09 dB of the orbicularis oris gap.Removing the largest conductivity contrast leaves all montage assignments unchanged and moves no gap across zero.Source: §3.3 · p. 10

3.4 Truncation sensitivity

  • Excluding jaw sites within 10 mm of the truncation face shifts the median gap by −0.17 dB, and no muscle changes sign.For muscles whose result does not move, the best jaw electrode was never a near-cut site.Source: §3.4 · p. 10

3.5 A homogeneous conductor reaches the same verdicts

  • A homogeneous soft-tissue conductor preserves every montage assignment, but eight of ten gap magnitudes shift beyond the 0.27 dB floor.The median shift is 0.482 dB and the maximum is 3.271 dB for mentalis; the detailed conductor is needed for decibel magnitudes, not montage verdicts.Source: §3.5 · pp. 10–11

3.6 Placement by anatomical target outperforms density

  • A target-chosen four-site ear cluster beats the median random four-site selection by 1.07 dB for lateral pterygoid and is equivalent for sternocleidomastoid.The respective lateral-pterygoid gaps are −1.564 and −0.498 dB; sternocleidomastoid gaps are −0.973 and −0.979 dB.Source: §3.6 · p. 11
  • For devices with few contacts, placement matters more than the number of candidate positions considered.The two sites that won the unmatched temporalis and sternocleidomastoid argmax were not in the target-chosen cluster.Source: §3.6 · p. 11

4. Discussion

The controlled comparison does not support complementary jaw and retroauricular montages: every resolved articulator favours the jaw, and the labial group is lost without a reliable muscle trade. The conclusion’s most model-dependent case is temporalis, whose verdict changes with the source-orientation treatment.

4.1 The montages are not complementary

  • Every resolved articulator favours the jaw; the three apparent ear-favouring muscles are nearest zero but do not cross it after orientation and electrode-count controls.Those muscles attach at or near the temporal bone.Source: §4.1 · p. 11
  • The retroauricular montage loses the labial group by 8.11 to 22.40 dB and returns nothing measurable in exchange.Anatomical proximity to a bony attachment predicts competitive sites, but not that those sites outperform the jaw.Source: §4.1 · p. 11
  • The no-ear-winner conclusion depends on temporalis fibre derivation: the uniform sweep favours the ear by 2.57 dB, while the derived-fibre interval crosses zero.The two intervals answer different questions; the authors use the derived field because it removes the uniform-orientation assumption.Source: §4.1 · p. 11

4.2 An anatomical prediction that did not survive

The anatomical prediction was that muscles attaching near the temporal bone would couple preferentially to retroauricular sites. Those muscles were the only ear-favouring cases in an uncontrolled comparison, but none remained so after matching electrode counts and deriving fibre directions.

4.2 An anatomical prediction that did not survive

  • The prediction targeted muscles near the temporal bone: temporalis originates beneath the superior cEEGrid row, sternocleidomastoid inserts on the mastoid, and lateral pterygoid inserts at the mandibular condyle.The condyle articulates with the temporal bone’s mandibular fossa.Source: §4.2 · p. 11
  • Only temporalis, sternocleidomastoid, and lateral pterygoid showed the predicted retroauricular advantage in the uncontrolled comparison; none survived matched counts and derived fibre directions.Source: §4.2 · p. 11
  • Proximity to a bony attachment predicts which sites are competitive, but does not establish that an electrode there observes the fibre better.The authors attribute the distinction to current returning through paths set by tissue conductivities, rather than proximity alone.Source: §4.2 · p. 11

4.3 The mechanism is distance, not intervening tissue

The jaw–ear coupling gap is set mainly by source-to-electrode geometry; tissue conductivity modifies it less and can shift it in either direction. The adipose contribution is especially small for the labial group, whose ear deficit is predominantly geometric.

4.3 The mechanism is distance, not intervening tissue

  • For the labial group, adipose–muscle conductivity contrast accounts for 0.6% of platysma’s gap and 13.3% of orbicularis oris’s; distance accounts for the remainder.The absolute material contribution is 0.03 to 2.86 dB, while the gaps it is measured against span an order of magnitude.Source: §4.3 · p. 12
  • −0.158 versus −0.045 dB/mm: retroauricular attenuation falls off 3.5 times faster than jaw attenuation over the shared 3–87 mm range.The fitted curves compare only the range where both montages have electrodes.Source: Figure 5 · p. 12
  • ρ = −0.955 (p = 0.001, n = 7): greater adipose path fraction accompanies a smaller material share because both routes shift together and cancel in the ratio.Leave-one-out correlations remain between −0.928 and −0.986, with p below 0.01 in each case.Source: §4.3 · p. 12, Figure 6 · p. 13
  • Removing the conductivity contrast enlarges temporalis’s gap from −3.801 to −4.923 dB, while the contrast supplies 21.0% and 17.4% of the sternocleidomastoid and lateral pterygoid gaps.The material term therefore opposes temporalis’s observed gap but acts with the gaps for the other two muscles.Source: §4.3 · p. 13
  • Geometry sets the jaw–ear gap, reaching 19.1 dB; the material term never exceeds 2.9 dB and shifts five muscles toward each montage.The material term changes sign across muscles, so a single material share for the set would obscure that variation.Source: §4.3 · p. 13, Figure 7 · p. 14
  • A smaller material share is predicted to transfer more readily between subjects; this prediction is unmeasured for temporalis, masseter, and medial pterygoid because they lack layer profiles.The stated share ranges from 0.6–13.3% across the labial group, versus 21% and 17% for sternocleidomastoid and lateral pterygoid.Source: §4.3 · p. 13

4.4 What this licenses for device design

With orientation and electrode count controlled, the retroauricular montage loses the labial group and has no reliable muscle advantage in return. The study does not identify an optimal montage, but supports anatomical targeting when choosing a small contact set.

4.4 What this licenses for device design

  • The retroauricular montage loses the labial group by 8.1–22.4 dB, and no articulator favours it after source orientation and electrode count are controlled.The authors describe this as a one-sided cost rather than a trade between muscle groups.Source: §4.4 · p. 14
  • Temporalis, sternocleidomastoid, and lateral pterygoid’s apparent unmatched ear advantages dissolve under matched four-electrode comparisons; their gaps depend on which four sites are carried.An unmatched comparison can make an advantage depend on site selection rather than anatomy.Source: §4.4 · p. 14
  • A four-site cluster chosen by anatomical target outperforms arbitrary four-site placement from the same candidates, leaving placement as the design lever for few-contact devices.The comparison is against the median of random four-site draws.Source: §4.4 · p. 14
  • The study does not identify an optimal montage because a best-four argmax across densely spaced candidates rewards candidate density rather than reliable placement.Ten of fourteen retroauricular positions lie on a cEEGrid path spaced 12–18 mm apart.Source: §4.4 · p. 15
  • Single-site argmax margins are only 0.03–1.02 dB, and four muscles’ margins are at or below the 0.65 dB meshing-floor bound; all ten runner-up sites remain in the same montage as the best site.Thus single-site winners are fragile even though montage-level assignments are not.Source: §4.4 · p. 15
  • Coupling strength is not decoding accuracy, and the model does not separate tongue or suprahyoid compartments, so its placement statements do not cover them.MIDA pools those compartments, which carry a large share of articulation.Source: §4.4 · p. 15

4.5 Contamination, described muscle by muscle

The model resolves retroauricular EMG contamination into muscle-specific contributions rather than treating it as one nuisance signal. That contact-specific description can also inform spatial filtering for rejection.

4.5 Contamination, described muscle by muscle

  • The retroauricular montage’s strongest relative couplings include temporalis, sternocleidomastoid, and lateral pterygoid, whose best positions differ; cg01 and cg08 therefore contain different muscle mixtures.The pooled suprahyoid compartment also contributes to the mixture, with its retroauricular field shown separately.Source: §4.5 · pp. 15–16
  • MIDA reports the suprahyoid sensitivity field through pooled Muscle (General), not as per-muscle values, because it does not segment the suprahyoid group.The figure overlays the mastoid notch, hyoid greater horn, and the corridor between them.Source: Figure 9 · p. 16
  • A spatial filter informed by which muscle dominates at each contact differs from one treating retroauricular EMG as a single nuisance component.The muscle-specific contamination description is usable for rejection as well as sensing.Source: §4.5 · p. 16

4.6 A specific prediction for a companion experiment

The paper predicts a null for a simultaneous, matched-channel experiment: no articulator should carry more information in retroauricular channels than in jaw channels. A reliable ear advantage for the named candidate muscles would falsify the model.

4.6 A specific prediction for a companion experiment

  • An eight-channel rig with four jaw and four retroauricular channels recording identical utterances should show no articulator for which ear channels carry more information, with the labial group degrading sharply at the ear.This is the paper’s stated prediction for a physical experiment recording both montages simultaneously.Source: §4.6 · p. 16
  • A reliable retroauricular advantage for temporalis, sternocleidomastoid, or lateral pterygoid would falsify the model.The authors suggest differing real-muscle fibre geometry or mechanical or acoustic coupling as likely explanations for such a result.Source: §4.6 · p. 16
  • Stating the null in advance makes a positive finding refuting evidence rather than an outcome explainable as insufficient sensitivity.The alternative prediction that the ear retains temporalis-driven gestures could be dismissed after a null result as lacking sensitivity.Source: §4.6 · p. 16

4.8 How the retroauricular advantage dissolved

Temporalis’ apparent retroauricular advantage weakened as orientation, electrode-count, and fibre-field assumptions were corrected, and the derived-field interval spans zero. The authors therefore do not treat the intermediate ear advantage as established.

4.8 How the retroauricular advantage dissolved

  • Temporalis’ apparent ear advantage shrank from −3.92 dB (field magnitude, best of 14 ear sites) to −2.57 dB (matched four-site counts; interval [−3.31, −0.03]), then derived fibres yielded −1.15 dB with [−1.45, +5.46] spanning zero.Source: §4.8 · pp. 16–17
  • The three corrections address orientation projection, unequal candidate-site counts, and assumed rather than anatomy-derived fibre directions.Source: §4.8 · p. 17
  • The final interval enumerates all 1,001 four-site retroauricular montages; 50.5 per cent favour the ear and 49.5 per cent favour the jaw.Source: §4.8 · p. 17
  • The monotone drift across corrections is suggestive but not probative; the pre-committed derived-field step crosses zero by removing an assumption.Source: §4.8 · p. 17
  • The intermediate result appeared credible because it cleared the measurement floor and matched the prediction that muscles near the temporal bone would favour ear sites.Source: §4.8 · p. 17

4.7 Limitations

The limitations constrain which muscles and anatomical configurations the model can test, and leave some verdicts less secure than the robust labial results. Several boundaries, including inter-subject variation and articulation-related geometry changes, remain unquantified.

4.7 Limitations

  • A same-resolution mesh rebuild moves gaps by up to 1.554 dB, wider than four reported margins; the five labial verdicts remain unaffected.Source: §4.7 · p. 17
  • Only ten of eighteen target muscles are modelled; posterior digastric and stylohyoid, which motivated the retroauricular hypothesis, are absent from the per-muscle comparison.Source: §4.7 · p. 17
  • A single MIDA anatomy cannot estimate between-subject variance, and the SCM and lateral pterygoid gaps are more dependent on tissue properties than geometry.Source: §4.7 · p. 17
  • The inferior boundary’s residual bias is unknown, while excluding the two near-cut jaw sites shifts the median gap by −0.17 dB and flips no signs.Source: §4.7 · pp. 17–18
  • The static model does not quantify conductor changes during articulation, when tongue, oral-cavity, and airway geometry move; the jaw montage is closer to those structures.Source: §4.7 · p. 18
  • The orientation treatment is uneven: a derived per-voxel field is available for some muscles, inappropriate for others, and impossible for at least one.Source: §4.7 · pp. 18–19

Tables

The tables document the model’s tissue inputs and site paths, quantify uncertainties, and report montage comparisons across robustness checks.

Tables

  • The tissue-conductivity table lists assigned values and sources for all 116 MIDA labels, while the layer-stack table reports target thickness and fat encountered before each target.Source: Tables · p. 19
  • The electrode-site tables give target and path measurements, and Supplementary Figure S1 inventories volume and depth for nine air-filled compartments.Source: Tables · pp. 19–20, Figure S1 · p. 22
  • The error budget distinguishes ratio-affecting terms from those that scale the whole lead field and cancel in reported comparisons.Source: Tables · p. 20
  • Tissue contrast affects gaps in different directions: −1.121 dB for temporalis, +0.411 dB for sternocleidomastoid, and +0.323 dB for lateral pterygoid.Source: Tables · p. 21
  • A same-resolution mesh rebuild moves per-muscle gaps by up to 1.554 dB; this is unquantified because one rebuild pair cannot establish a distribution.Source: Tables · p. 21
  • The five labial muscles show robust jaw-favouring gaps, from +8.11 to +22.40 dB, with all five jaw subsets and 100.0% orientation agreement.Source: Tables · p. 21

Data and code availability

The analysis materials are available online, while the MIDA model must be obtained separately under its licence.

Data and code availability

Loading 2610.03723v1…