# Fileset

[Adv Elect Materials - 2026 - Nishioka - IonGating Reservoir Computing for Preprocessing Free Speech Recognition_のコピー.pdf](https://mdr.nims.go.jp/filesets/a4779bd3-4b19-49e9-b57f-526df349a743/download)

## Creator

[Daiki Nishioka](https://orcid.org/0000-0002-3369-7700), [Akito Tateyama](https://orcid.org/0000-0002-0403-3462), [Hina Kitano](https://orcid.org/0009-0008-9132-0275), [Takashi Nakanishi](https://orcid.org/0000-0002-8744-782X), [Kazuya Terabe](https://orcid.org/0000-0003-3988-3456), [Takashi Tsuchiya](https://orcid.org/0000-0002-6950-6160)

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Ion‐Gating Reservoir Computing for Preprocessing‐Free Speech Recognition from Throat Vibrations](https://mdr.nims.go.jp/datasets/1b870ea2-3e17-4d58-9534-7b4ad9abb6f4)

## Fulltext

Advanced Electronic Materials www.advelectronicmat.deRESEARCH ARTICLEIon-Gating Reservoir Computing for Preprocessing-Free Speech Recognition from Throat Vibrations Daiki Nishioka1 Akito Tateyama2 , 3 Hina Kitano2 , 4 Takashi Nakanishi2 , 3 Kazuya Terabe2 Takashi Tsuchiya2 , 4 1 International Center for Young Scientists (ICYS), National Institute for Materials Science (NIMS), Tsukuba, Ibaraki, Japan 2 Research Center for Materials Nanoarchitectonics (MANA), National Institute for Materials Science (NIMS), Tsukuba, Ibaraki, Japan 3 Division of Soft Matter, Graduate School of Life Science, Hokkaido University, Sapporo, Japan 4 Department of Applied Physics, Tokyo University of Science, Tokyo, Japan Correspondence: Daiki Nishioka ( NISHIOKA.Daiki@nims.go.jp) Takashi Nakanishi ( NAKANISHI.Takashi@nims.go.jp) T akashi Tsuchiya ( TSUCHIYA.Takashi@nims.go.jp) Received: 12 January 2026 Revised: 24 February 2026 Accepted: 21 April 2026 Keywords: edge artificial intelligence | gel-electret mechanoelectric generator | graphene electric-double-layer transistor | ion-gating reservoir | physical reservoir computing | spoken-digit recognition | throat-vibration sensing ABSTRACT Physical reservoir computing offers a promising route to efficient edge-AI, yet most reservoir-based speech-recognition demonstrations still rely on microphones and digital preprocessing. Here, we report a real-time spoken-digit recognition system that directly processes throat vibrations by integrating a π–gel-electret mechano-electric generator (MEG) sensor with a multi- channel ion-gel/graphene ion-gating reservoir (IGR). The MEG, a free-deformable electret-based vibration sensor, converts articulation-induced throat-surface vibrations into voltage signals, which are then transformed into pulse sequences and applied directly to the IGR gate without any spectral or handcrafted feature extraction. Channel-dependent Dirac-point shifts, nonlinear electric-double-layer dynamics, and distinct relaxation behaviors in the IGR expand the one-dimensional MEG output into high- dimensional reservoir states, which are decoded using a simple linear classifier. This hybrid MEG–IGR system achieves an impressive real-time spoken-digit recognition accuracy of 96.8%, demonstrating that ion-gating reservoirs can efficiently extract discriminative spatiotemporal features from raw biomechanical signals. By eliminating frequency-domain preprocessing and enabling compact, low-power front-end computation, the approach provides a viable pathway for on-device speech interfaces. Beyond speech recognition, direct throat-vibration sensing opens opportunities for robust human–machine interfaces, assistive communication technologies, and secure authentication systems based on spoof-resistant biomechanical signatures. 1S  m  o  a  t  fs           Tc©Ah Introduction peech recognition has become a fundamental component ofodern human–machine interfaces, enabling voice-controlledperation in mobile devices, wearables, smart appliances, andssistive technologies [ 1 ]. Conventional speech-recognition sys-ems rely on microphones to capture airborne acoustic signals,ollowed by computationally intensive digital preprocessing—uch as Fourier-based spectral decomposition, mel-frequencyhis is an open access article under the terms of the Creative Commons Attribution License, which permited. 2026 The Author(s). Advanced Electronic Materials published by Wiley-VCH GmbH dvanced Electronic Materials , 2026; 12:e00006 ttps://doi.org/10.1002/aelm.202600006cepstral coefficient (MFCC) extraction, and deep neural net-work (DNN) inference [ 2–4 ]. While these methods have ledto remarkable advances in accuracy, they remain vulnerable toenvironmental noise [ 5 ] and require considerable power andcomputational resources [ 6 ], making them unsuitable for low-power edge devices. Moreover, recent advances in generative AIhave made it increasingly easy to synthesize natural-soundingspeech, raising concerns about the security and authenticityof microphone-based voice interfaces [ 7, 8 ]. An alternativeits use, distribution and reproduction in any medium, provided the original work is properly 1 of 12http://www.advelectronicmat.dehttps://doi.org/10.1002/aelm.202600006https://orcid.org/0000-0002-3369-7700https://orcid.org/0000-0002-8744-782Xhttps://orcid.org/0000-0003-3988-3456https://orcid.org/0000-0002-6950-6160mailto:NISHIOKA.Daiki@nims.go.jpmailto:NAKANISHI.Takashi@nims.go.jpmailto:TSUCHIYA.Takashi@nims.go.jphttp://creativecommons.org/licenses/by/4.0/https://doi.org/10.1002/aelm.202600006http://crossmark.crossref.org/dialog/?doi=10.1002%2Faelm.202600006&domain=pdf&date_stamp=2026-04-26a  c  v  Rd  d  p  p  s  b  i  p  i  p  t  e  r  r  r  s  s  s  s  tP  a  B  o  e  t  A  v  o  e  c  [  s  m  m  f  t  (  t  c  t  l  a  i  c  f  t  s  a  e  c  t  a  r  s  s                                                     2 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creapproach to speech recognition is to measure the mechani-al vibrations of the throat resulting from the motion of theocal folds and surrounding tissues during articulation [ 9–14 ].ecent wearable silent-speech interfaces and “artificial throat”evices have expanded this direction by employing diverse trans-uction mechanisms—including strain-/iontronic sensors andhotonic implementations (e.g., optical-fiber–based wearablehotonic artificial throats)—to decode speech content and enableilent communication from neck-surface motions [ 15 ]. Moreroadly, rapid progress in photonic biointerfaces—spanning flex-ble optical-fiber platforms, wearable photonic wristbands forulse-wave–based cardiorespiratory monitoring and biometricdentification, and chip-scale photonic sensors for cuffless blood-ressure measurements—highlights the expanding photonicoolbox for robust, miniaturizable physiological sensing at thedge [ 16–18 ]. Such biomechanical vibration signals are inherentlyobust against environmental acoustic noise because they do notely on airborne sound propagation [ 9, 12, 14 ]. Furthermore, theyeflect physiological characteristics that are extremely difficult topoof, offering unique opportunities for secure and noise-resilientpeech interfaces. However, even when vibration or contactensors are employed, existing systems still depend on digitalignal processing or deep-learning pipelines [ 9–14 ], which limitheir suitability for compact, low-power, real-time applications. hysical reservoir computing (PRC) has recently emerged as promising paradigm for efficient edge-AI systems [ 19, 20 ].y harnessing the intrinsic dynamics and nonlinear responsesf physical devices, PRC enables high-dimensional featurexpansion and temporal processing with minimal computa-ional overhead, requiring only a linear readout for training.mong various physical reservoirs [ 21–40 ], ion-gating reser-oirs (IGRs)—which exploit ionic gating–induced modulationf electronic material properties—are particularly attractive fordge-AI implementations due to their flexible material design,ompact device structures, and strong computational capability 41–54 ]. For integrating PRC directly with sensors to con-truct fully physical edge-AI systems, however, the reservoirust additionally accommodate wide sensor bandwidths andaintain high computational efficiency to achieve strong per-ormance with relatively small node counts. In this context,he ion-gel/monolayer-graphene electric-double-layer transistorEDLT)–based IGR that we recently developed achieves excep-ionally high performance, broad-band operation, and excellentomputational efficiency [ 53 ]. These capabilities originate fromhe ambipolar transport of graphene, which provides rich non-inear responses, combined with multiple relaxation processesrising from ionic motion in the gel electrolyte and the charg-ng/discharging dynamics of the electric double layer (EDL). Thisombination yields a highly capable physical reservoir suitableor real-time processing of time-varying sensor signals. Despitehese advances, demonstrations of reservoir-computing–basedpeech recognition have largely focused on microphone-acquiredudio and typically rely on spectral or handcrafted featurextraction [ 21–28 ]. In contrast, the direct processing of biome-hanical throat-vibration signals—without frequency-domainransformations or feature-engineering steps—has received rel-tively little attention. Furthermore, the behavior of ion-gatingeservoirs when driven by such vibration signals and theiruitability for real-time edge-AI implementations have yet to beystematically explored. of 12Here, we address these challenges by integrating a π–gel-electretmechano-electric generator (MEG) sensor [ 55 ] with a multi-channel ion- gel/graphene IGR [ 53 ] to realize preprocessing-free,real-time spoken-digit recognition. The π–gel- electret is a vis-coelastic gel material made by converting π-conjugated molecules(e.g., porphyrin [ 56 ], fullerene [ 57 ], pyrene [ 55 ], conjugatedpolymer [ 58 ], etc.) substituted with branched alkyl chains into asolvent-free liquid at room temperature, gelating them with theaddition of a low-molecular-weight gelator, and then electretizingthem through corona charging. The MEG is a free-deformableelectret-based vibration sensor in which a corona-poled π–gelretains quasi-permanent charges, enabling sensitive detectionof throat-surface vibrations through deformation-induced poten-tial changes while remaining intrinsically robust to ambientacoustic noise. The resulting one-dimensional voltage stream isconverted into pulse-voltage sequences and applied directly tothe IGR gate. Through channel-dependent Dirac-point shifts,nonlinear EDL dynamics, and distinct relaxation behaviors, theIGR transforms this input into eight diverse current responses,which are further expanded into a high-dimensional reservoirstate and decoded by a linear readout. Despite the MEG’s stan-dalone recognition accuracy being only 18.2%, using this hybridMEG–IGR reservoir, we achieve real-time spoken-digit recogni-tion accuracy of 96.8%, demonstrating that the IGR effectivelyextracts discriminative spatiotemporal features embedded in rawbiomechanical vibration signals. By eliminating the need forspectral analysis, hand-crafted features, or deep-learning-basedpreprocessing, our approach enables a compact and low-powerphysical-computing architecture suitable for on-device speechrecognition. Taken together, this study establishes an entirelyphysical, preprocessing-free reservoir-computing framework forreal-time speech recognition. Beyond speech interfaces, thedemonstrated framework highlights the potential of throat-vibration sensing and ion-gating reservoirs for wearable com-munication devices, assistive technologies for individuals withspeech impairments, and secure authentication systems based onbiomechanical signatures. 2 Results and Discussion 2.1 Concept and Task Design for MEG–IGR–Based Speech Recognition Figure 1a illustrates the conceptual framework of the speech-recognition system developed in this study, which integrates anMEG–based vibration sensor with an IGR. Instead of a conven-tional microphone, a π–gel- electret MEG device [ 55 ] was usedto detect vibrations around the throat (the Adam’s apple) duringarticulation. The MEG converts these mechanical vibrations intovoltage signals. These raw voltage traces were transformed intopulse-voltage waveforms at the sensor’s original sampling rate,and the resulting pulses were applied directly to the gate terminalof the IGR [ 53 ]. The corresponding gate, source, and draincurrent responses were regarded as reservoir states and fed to alinear readout layer to classify spoken digits. Because the IGRnonlinearly expands the MEG signals into a high-dimensionalspatiotemporal representation, it enables accurate speech clas-sification using only a lightweight linear model. Importantly,this framework eliminates the need for computationally inten-sive preprocessing such as Fourier-based spectral analysis orAdvanced Electronic Materials, 2026tive Commons LicenseFIGURE 1 Concept and data acquisition for MEG–IGR–based speech recognition. (a) Schematic illustration of the proposed speech-recognition framework integrating a π–gel-electret MEG vibration sensor with an IGR. MEG output voltages are converted into pulse signals and applied to the IGR, whose current responses form reservoir states for linear readout. (b) Potential applications of the MEG–IGR approach, including noise-resilient interfaces, assistive systems for individuals with speech impairments, and secure authentication based on throat-vibration patterns. (c) Placement of the MEG sensor on the throat, slightly below the laryngeal prominence, for acquiring vibration signals during digit articulation. (d) Representative MEG waveforms corresponding to spoken digits “zer o” thr ough “nine,” demonstrating distinct temporal vibration patterns associated with each utterance. h  p  i  v  a  [  m  i  t  [r  aT  t  b  t                   A 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaandcrafted feature extraction, highlighting its potential for low-ower, edge-AI applications. The anticipated application spaces outlined in Figure 1b . Since the method relies on throat-ibration sensing rather than airborne sound, it naturally avoidscoustic coupling and is less susceptible to environmental noise 9, 12, 14 ]. Moreover, the ability to track biomechanical throatotion offers potential use for assistive interfaces supportingndividuals with speech impairments [ 14 ]. As generative AIechnologies increasingly enable the spoofing of audible speech 7, 8 ], the biomechanical nature of throat-vibration patterns—ooted in vocal-fold dynamics and tissue motion—also suggestspplicability for secure personal identification systems. o establish a dataset for evaluating the proposed framework,he MEG sensor was attached to the speaker’s throat, slightlyelow the laryngeal prominence, as shown in Figure 1c . Aotal of 500 samples were collected by recording 50 repetitionsdvanced Electronic Materials, 2026each of the spoken digits “zer o” thr ough “nine.” Repr esentativeMEG outputs for the ten digits are presented in Figure 1d .Although these waveforms differ fundamentally from acousticspeech signals—because they arise from mechanical vibrationsof the throat surface—they nevertheless exhibit clear digit-dependent temporal patterns. The MEG used in this work isa deformable π–gel- electret–based mechanoelectric generator,constructed by infiltrating an alkyl-pyrene liquid gelated witha low–molecular-weight gelator into a polyurethane nonwovenmatrix and subsequently corona-poling the structure to retainquasi-permanent charges [ 55 ]. Mechanical deformation of theelectret layer modulates its internal and surface charge distri-bution, thereby generating voltage fluctuations proportional tothe applied vibration. Owing to its electret-based transductionmechanism, the MEG sensor can detect subtle biomechanicalsignatures associated with articulation, providing rich temporalinformation that is highly discriminative across spoken-digit3 of 12tive Commons LicenseFIGURE 2 Electrical characteristics of the ion-gel/graphene EDLT-based IGR. (a) Schematic illustration of the ion-gel/graphene IGR and optical micrograph of the multi-channel device comprising mono-layer graphene channels with different lengths (5 mm, 20 µm, 100 µm) and widths (30 and 80 µm), all sharing a common gate electrode. (b–g) Pulsed transfer characteristics of each graphene channel measured under distinct drain biases ( VD1 = –1 V, VD2 = –0.5 V, VD3 = –0.1 V, VD4 = + 0.1 V, VD5 = + 0.5 V, VD6 = + 1 V). Gate and drain voltages were applied as 10 ms pulses with a 1 s interval. The inset of (e) shows the Dirac cone of graphene, where modulation of the Fermi level EF relative to the Dirac point by the applied VG results in p -type or n -type behavior. c  p  r  r  r  p  c  i  t  w  e  d  t2GMF  p  o  dl  8  p  g  t  E  i  c                                 4 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Crelasses. Details of the fabrication procedure and material com-osition can be found in the Methods section and the prioreport [ 55 ]. Despite the presence of such discriminative tempo-al features, conventional speech-recognition pipelines typicallyequire substantial digital processing—including spectral decom-osition, feature extraction, and deep neural inference—posinghallenges for low-power edge implementation. In this study, wenstead exploit the intrinsic nonlinear dynamics and computa-ional richness of the IGR to process raw MEG outputs directly,ithout any explicit preprocessing. This approach allows us tovaluate the feasibility of an edge-compatible speech-recognitionevice in which a physical reservoir extracts and transforms theemporal structure embedded in throat-vibration signals. .2 Electrical Characteristics of the raphene-Based IGR and Its Nonlinear Mapping of EG Outputs igure 2a illustrates the ion-gel/graphene IGR used as thehysical reservoir layer in this study [ 53 ]. As shown in theptical micrograph inset, the device comprises multiple indepen-ent single-layer graphene channels with different geometries—engths of 5 mm, 20 µm, and 100 µm, and widths of 30 and0 µm—sharing a common gate electrode. Six drain terminalsrovide six independent drain-current ( ID ) responses to a singleate input. When a gate voltage ( VG ) is applied, mobile ions withinhe electrolyte migrate toward the graphene interface and form anDL [ 49, 59–61 ], leading to highly efficient electrostatic carriernjection into graphene. The resulting modulation of channelonductance is detected as the drain current. Further detailsof 12aof the device physics underlying this ion-gel/graphene IGR canbe found in our previous studies [ 53 ]. Figure 2b–g shows thetransport characteristics of each channel. To isolate the EDL-driven carrier injection response from slow dynamical processesoften observed in graphene transistors [ 62, 63 ], both the gateand drain voltages were applied as pulses, and the drain-currentvalues at each pulse were plotted. With a pulse width of 10 msand an interval of 1 s, all channels exhibit characteristic V-shapedtransfer curves associated with the linear dispersion of monolayergraphene. Such strong nonlinear transfer characteristics directlycontribute to the nonlinear mapping capability essential forreservoir computing. To further diversify the reservoir states, eachdrain terminal was biased at a different voltage ( VD1 = –1 V, VD2 =–0.5 V, VD3 = –0.1 V, VD4 = + 0.1 V, VD5 = + 0.5 V, VD6 = + 1 V).As evident in Figure 2b–g , the Dirac point (DP) systematicallyshifts with VD , originating from the decrease in the effectivegate voltage, VG,eff = VG − VD /2. This channel-dependent DPshift provides multiple distinct nonlinear transfer functions thatenrich the high-dimensional reservoir state space. In addition toDP shifts, the magnitude of VD also affects the high- VG plateauregion. In graphene FETs, this plateau commonly arises frommobility degradation under high fields due to carrier scattering[ 64 ]. In ion-gated transistors, an additional effect emerges: pulsedgate voltages induce EDL charging/discharging currents thatcontribute to the measured gate current, and when the draincurrent is comparable in magnitude, a portion of this ioniccurrent is superimposed on the drain current [ 41, 53 ]. This effectis particularly visible in high-resistance channels under smalldrain biases—for example, ch3 at VD = –0.1 V. Under theseconditions, relatively large negative gate pulses (approximately–1 to –2 V) generate gate currents that are no longer negligibleAdvanced Electronic Materials, 2026tive Commons LicenseFIGURE 3 IGR processing of MEG vibration signals and class separability. (a–c) Pulse-shaped VG sequences converted from MEG outputs corresponding to the spoken digits one, four, and nine (upper panels), and the resulting IGR current responses (lower panels). The inset in (a) magnifies the gate-current spikes induced by EDL charging/discharging and the relaxation-type drain-current dynamics. (d) Conventional non-real-time LDA of the full time-evolution of the IGR current responses for 500 spoken-digit samples. The first two discriminant axes (LD1, LD2) yield apparent clustering of the ten digit classes. (e) LDA applied directly to the raw MEG sensor outputs, showing insufficient class separation. Comparison with (d) demonstrates that the IGR nonlinearly expands the MEG signal into a high-dimensional feature space, enabling linear discrimination. c  t  s  c  o  t  rF  o            A 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaompared with the drain current. As a result, as indicated byhe purple arrow in Figure 2d , the drain current exhibits atrong nonlinear response that resembles an apparent increase inhannel resistance as VG becomes more negative. The coexistencef these device-state-dependent nonlinear dynamics indicateshat the ion-gel/graphene IGR provides a rich nonlinear basis foreservoir computing [ 53 ]. igure 3a–c presents the pulse-shaped gate-voltage sequencesbtained from MEG outputs for the spoken digits one, four, anddvanced Electronic Materials, 2026nine (upper panels), along with the corresponding IGR currentresponses (lower panels). The MEG sensor operated at 500 Hz,and the IGR was driven synchronously with 500 Hz VG pulses(50% duty cycle). The pulse amplitude during the on-state wasset to 20 times the MEG voltage (scaled to a maximum of 4 V),while the off-state was fixed at 0 V. During these measurements,six drain currents ( ID1–ID6 ) and the gate and source currents ( IGand IS ) were recorded. Thus, the IGR expands a one-dimensionalMEG input into eight-dimensional current responses that serveas the reservoir states. As seen in Figure 3a–c , the eight current5 of 12tive Commons Licenser  s  o  l  w  B  l  i  s  a  c  s  n  dT  i  s  s  s  s  f  s  e  w  [  s  a  s  d  a  c  s  L  s  t  t  m  r  t  s  s  o  u  r  [  l  r  a  r  r  2PF  n  T  p  a                                                          6 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaesponses exhibit distinct temporal patterns depending on thepoken digit. The inset of Figure 3a shows a magnified regionf the input–output relationship: the gate current displays spike-ike features associated with EDL charging and discharging [ 41 ],hereas the drain currents follow relaxation-type dynamics.ecause each channel experiences a distinct Dirac point and non-inear transfer characteristic due to its drain bias, the same gatenput produces diverse current responses across channels. Undermall drain biases (e.g., ID3 and ID4 ), gate-current-induced spikesppear in the drain current due to the relatively large ionic currentontribution. Whether such spikes emerge depends on devicetate and bias conditions, introducing additional state-dependentonlinearities that further enrich the reservoir’s computationaliversity. o use these current responses for real-time speech recognition,t is necessary to sequentially extract temporal features (reservoirtates) from them and feed these features into a linear clas-ifier. In this way, the linear classifier can output a temporalequence of decisions. Although the details of this real-timepeech-recognition scheme are discussed in the next section, weirst examine the characteristics of the IGR current responseshown in Figure 3a–c —particularly their waveform shapes—andvaluate the latent separability among classes. For this purpose,e performed a non–real-time linear discriminant analysis (LDA) 65 ] using the IGR current responses corresponding to all 500peech samples. For each speech sample, all temporal pointscross the eight current responses were treated as features. Ashown in Figure 3d,e , the dimensionality was reduced to twoiscriminant axes (LD1 and LD2) that maximize the separationmong the ten digit classes. In Figure 3d , the full temporalurrent-response features generated by the IGR form clearlyeparated clusters corresponding to each digit. In contrast, whenDA is applied directly to the raw MEG signals (Figure 3e ), theeparability is significantly degraded. These results indicate thathe IGR current responses shown in Figure 3a–c nonlinearlyransform the raw MEG outputs into a representation that iseaningful for class discrimination. Furthermore, classes that lieelatively close to one another in this discriminant space tendo be more difficult to classify, consistent with their waveformimilarity. However, this LDA analysis utilizes the full timeequence of each response, making it unsuitable for real-timeperation because classification can only occur after the entiretterance is complete. Conventional speech-recognition systemsely on sequential preprocessing (e.g., MFCCs or cochleagrams) 21–28 ], but these impose a substantial computational burden thatimits deployment on edge devices. To address this, we feed theaw MEG signals directly into the IGR without any preprocessingnd sequentially sample the resulting current responses, enablingeal-time reservoir-state generation. This real-time spoken-digitecognition strategy is described in detail in the following section..3 Spoken-Digit Recognition Using MEG–IGR hysical Reservoir Computing igure 4a illustrates the framework of the spoken-digit recog-ition task implemented using the MEG sensor and the IGR.he voltage output from the MEG was converted into a 500 Hzulse-voltage sequence—matching the MEG sampling rate—andpplied directly to the gate terminal of the IGR. From the resultingof 12eight-dimensional current responses (examples in Figure 3a–c ), aset of NV virtual nodes was sequentially sampled within a timewindow tW . The resulting reservoir size N was therefore ( N =8 NV ), and the interval between successive nodes was θ. For eachtime window, the reservoir-state vector x = ( x1 , x2 ,. . . , xN , 1) waslinearly transformed using a readout weight matrix W to generatea ten-dimensional temporal output y = ( y0 , y1 ,. . . , y9 ) as follows: 𝒚 = 𝑊𝒙 (1)The readout weights W were trained using ridge regression withone-hot encoded target vectors for digits 0–9, as described inthe Methods section. For classification, the temporal outputsobtained across all time windows within a single sample weresummed to form an integrated output vector 𝒚̃ . The predicteddigit was then determined using a winner-take-all (WTA) rule[ 21–28 ], in which the index corresponding to the maximumcomponent of the integrated vector was selected: Answer =argma x𝑖 (𝑦𝑖 ) . Figure 4b,c summarizes the dependence of the test accuracyon the tW and node interval θ, obtained from five-fold cross-validation (400 training and 100 test samples). Under the condi-tion θ = 500 µs (Figure 4b ), the MEG–IGR system achieved 84.2%accuracy at tW = 2 ms, which increased with tW and saturatedat approximately 95% around tW = 16 ms. This improvementstems from the fact that a larger tW enables the reservoir tocapture temporal correlations over longer time scales, effectivelyenhancing its memory. However, increasing tW reduces real-time responsiveness and—under fixed θ—increases the numberof virtual nodes, leading to higher computational cost. Thecondition indicated by the black arrow in Figure 4b correspondsto tW = 2 ms, where four virtual nodes were extracted perpulse (i.e., N = 32). The corresponding confusion matrix andtest accuracy are shown in Figure 4d . Remarkably, despite theextremely small reservoir size, the MEG–IGR system achieved84.2% accuracy, outperforming the MEG-only baseline shown inupper panel of Figure 4e . The MEG-only result was obtainedby directly feeding an equivalent number of points (32 pointscorresponding to tW = 64 ms) into a linear classifier, effectivelytreating the MEG sensor as a physical reservoir. The low accuracyof the MEG-only baseline ( ≈ 18%; Figure 4e , upper panel) can berationalized by the LDA result in Figure 3e , which indicates thatthe raw throat-vibration dynamics measured by the MEG do notprovide a representation that is readily linearly separable acrossspoken-digit classes. Consequently, approaches that directly feedthe sensor output into a linear model have limited discriminativepower for this task. Importantly, however, the MEG waveformsstill embed digit-dependent information in a form that can beexploited by nonlinear processing or suitable feature extraction.Consistent with this, a conventional digital baseline that extractsmel-frequency cepstral coefficients (MFCCs) from the MEGsignals and applies a linear classifier achieves a test accuracyof 91.6% (Figure 4e , lower panel). Details of the MFCC-basedpipeline are provided in Note S1 and Figure S1 . These comparisonshighlight that the IGR functions as a compact physical feature-expansion front end that renders the MEG signal more linearlydecodable, while avoiding conventional digital preprocessing.The substantial improvement achieved with the IGR confirms itsability to extract and amplify discriminative features embedded inthe MEG signal. This arises from the high information-processingAdvanced Electronic Materials, 2026tive Commons LicenseFIGURE 4 Real-time spoken-digit recognition using the MEG–IGR reservoir. (a) Schematic of the real-time spoken-digit recognition framework. MEG outputs are converted into 500 Hz pulse-voltage sequences and applied to the IGR gate. From the resulting eight-dimensional current responses, NV virtual nodes are sampled within a time window tW with interval θ, forming a reservoir state vector used for linear readout. (b) Dependence of test accuracy on tW at a fixed node interval θ = 500 µs. The black arrow marks the minimal-node configuration ( tW = 2 ms, N = 32). (c) Dependence of test accuracy (left axis) and virtual node count NV (right axis) on θ at a fixed tW = 16 ms. Reducing θ increases temporal resolution and boosts performance, with the smallest θ = 27.8 µs yielding the highest accuracy. (d) Confusion matrix for the minimal-node condition ( tW = 2 ms, N = 32), showing 84.2% test accuracy. (e) Baseline results using the MEG signal without the IGR. Upper: MEG-only PRC baseline obtained by feeding 32 MEG samples directly into a linear classifier. exhibiting significantly lower accuracy than the MEG–IGR system. Lower: Conventional digital baseline in which mel-frequency cepstral coefficients (MFCCs) are extracted from the MEG signals and classified using a linear model. (f) Confusion matrix for the best-performing condition ( tW = 16 ms, θ = 27.8 µs, NV = 576, N = 4608), achieving 99.8% training accuracy and 96.8% test accuracy. Advanced Electronic Materials, 2026 7 of 12 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensec  m  [  s  d  t  p  m  g  t  s  c  t  e  cF  a  f  p  a  n  a  d  T  µ  t  t  r  i  T  cI  u  r  c  i  A  (  o  a  w  n  t  c  d  r  r  u  F  n  o  e w  c  f  f  t  c  a  t                                                      8 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaapacity (IPC) [ 66 ] —including nonlinear transformation andemory—and the intrinsic high dimensionality of the IGR 53 ]. By projecting the input into a space spanned by pasttates and their higher-order interactions, the IGR enables lineariscrimination of spoken-digit patterns. Further details regardinghe IPC characteristics of this IGR platform are provided in ourrevious studies. The confusion matrix in Figure 4d reveals thatisclassifications occur predominantly among neighboring digitroups (e.g., 0–3 and 6–9). This is consistent with Figure 3d , wherehese digit classes occupy nearby positions in the discriminantubspace. Improved separability among such closely spacedlasses can be achieved by increasing the temporal resolution ofhe reservoir states—i.e., decreasing θ and increasing NV—whichnables finer sampling of the temporal dynamics of the IGRurrents. igure 4c plots the dependence of classification accuracy (leftxis) and the number of virtual nodes NV (right axis) on θ at aixed time window of tW = 16 ms (corresponding to eight MEGulses). Across all θ, accuracies above 94% were consistentlychieved. As expected from NV = tW / θ, reducing θ increases theumber of nodes and improves classification performance byllowing the reservoir to capture the spike-like and relaxationynamics of the IGR output with higher temporal precision.he highest performance was obtained at the smallest θ = 27.8s, indicated by the black arrow in Figure 4c , yielding 99.8%raining accuracy and 96.8% test accuracy (Figure 4f ). Notably,his test accuracy slightly exceeds the 96.2% accuracy recentlyeported for analog in-memory computing using a memristor-mplemented deep neural network that extracts features from theI-46 spoken-digit dataset without digital preprocessing [ 67 ]. Inontrast to such multilayer nonlinear architectures, the MEG–GR system achieves comparable or even superior performancesing only a single physical reservoir layer followed by a lineareadout. This highlights the strong nonlinear transformationapabilities of the ion-gating reservoir and its efficiency in extract-ng discriminative features directly from raw sensor signals.lthough this condition yields a relatively large reservoir size NV = 576, N = 4608), Figure 4c demonstrates a clear trade-ff between reservoir size and computational performance: bydjusting θ, the accuracy can be tuned between 94.4% and 96.8%hile varying NV from 32 to 576. Here, the number of physicalodes NP was fixed at eight, and the overall reservoir size wasuned solely by modulating NV through θ. This adjustmentorresponds to controlling how extensively the intrinsic highimensionality arising from the time evolution of the IGR currentesponses is exploited for information processing. Figure S2aeproduces the NV dependence of the classification accuracynder the fixed physical node condition ( NP = 8; equivalent toigure 4c ), whereas Figure S2b shows the dependence on theumber of physical nodes when NV is fixed at 576. Comparisonf these trends indicates that the number of physical nodesxerts a larger influence on the overall performance than NVithin the investigated range. This suggests that the dominantontribution to the reservoir’s effective dimensionality originatesrom the diversity of the physical channels, arising not onlyrom geometry induced variations in channel resistance (andhus relaxation times), but also from the diversity in transportharacteristics (i.e., nonlinearity) introduced by independentlypplying distinct VD biases. Consequently, modulating NV (i.e.,uning θ) provides a practical knob to flexibly balance com-of 12putational cost and performance while maintaining a highaccuracy above 94%. This adaptive tunability between reservoir size and classifica-tion accuracy is highly advantageous for edge-AI deployment.Depending on system requirements—desired accuracy, permissi-ble power consumption, available computational resources, andwhether training is performed on-device or in the cloud—theMEG–IGR reservoir can be flexibly configured to operate in eithera low-resource or high-performance mode. Such configurability isessential for practical implementations of edge-intelligent sens-ing platforms. In addition, preliminary power characterization(Note S2 and Table S1 ) indicates that the IGR consumes anaverage electrical power of 5.09 mW under the present pulsed-bias conditions, which corresponds to an energy cost of 3.72mJ per word inference (i.e., the time-integrated power over oneutterance). Because the overall energy budget scales with boththe applied drain bias and the number of sampled reservoirstates, a low-power operating mode can be implemented bysimultaneously reducing the drain voltages to ∣ VD ∣≤ 0.1 V anddecreasing NV via coarser time multiplexing (larger θ, e.g., θ= 500µs). Under this combined setting, the average power is expected todecrease to approximately 150 µW, corresponding to ∼ 109 µJ perword. Although the reservoir states and the readout outputs aregenerated sequentially in real time, the final classification inthe present implementation is obtained by applying a WTAdecision to the time ‑integrated output over the entire utterance.Consequently, an end ‑to ‑end decision latency arises until a wordis fully articulated. In our dataset, the average waveform durationis 0.73 s; therefore, if the decision is made after integratingthe readout over the full waveform, the latency is on theorder of this duration (plus negligible computation time for thelinear readout). To explore latency reduction, we evaluated anearly ‑decision strategy in which the WTA decision is computedusing only the initial portion of each utterance. Figure S3 plots theclassification accuracy as a function of the decision time (i.e., thelength of the initial segment used for WTA). While the accuracyis low immediately after speech onset, it increases rapidly andreaches approximately 83% at around 200 ms. This result indicatesthat the MEG–IGR system can perform fast classification beforethe utterance ends by leveraging early spatiotemporal features,providing a viable route to low ‑latency operation. To put theseresults in a broader context, Table S2 provides a side-by-side com-parison with representative PRC approaches, summarizing keypractical metrics including classification accuracy, preprocessingrequirements, power/energy consumption, and decision latency.Overall, the MEG–IGR platform offers a competitive and highlyconfigurable operating space—ranging from low-resource, low-power settings to higher-performance modes—while retainingthe ability to extract discriminative features directly from rawsensor signals. 2.4 Outlook In practical operating environments, however, the classificationtask is expected to become more challenging due to noiseand motion ‑induced artifacts (e.g., sensor displacement, skinmotion, and mechanically induced baseline fluctuations). In suchAdvanced Electronic Materials, 2026tive Commons Licenses  a  d  r  m  t  u  fA  b  l  n  s  f  i  t  b  s  p  o  p  t  s  i  cA  g  a  i  a  a  e  m  t  p  p  d  f  b  v3I  n  M  c  a  r  o  h  I  p  i  s  d  c  p                                                A 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaituations, one of the most effective approaches is to incorporate,s training data, as many samples as possible acquired underiverse noise levels and motion conditions. By exposing theeadout to such variability during training, the system can learnore invariant decision boundaries and become more robusto real ‑world perturbations. Systematic validation of robustnessnder controlled noise and motion scenarios will be pursued asuture work. s further future directions, it is important to assess scalabilityeyond isolated spoken digits, including extension to differentanguages, larger vocabularies, and continuous ‑speech recog-ition. Notably, in our framework the MEG and IGR operateynchronously, and the readout produces outputs sequentiallyor each time window, enabling inherently streaming (real ‑time)nference. This property is expected to be advantageous for con-inuous recognition, where predictions can be updated windowy window and combined with temporal segmentation/decodingtrategies. At the same time, large ‑vocabulary natural ‑languagerocessing (e.g., language modeling) typically exceeds the scopef conventional time ‑series tasks addressed by reservoir com-uting alone. In such cases, the MEG–IGR device is anticipatedo function as a low ‑power front ‑end that extracts informativepatiotemporal features from raw biomechanical signals andnterfaces with higher ‑level digital processors dedicated to moreomplex linguistic inference. nother promising direction is on ‑device learning. By inte-rating the IGR with a high ‑performance memristive array asn in ‑memory readout network, it may become possible tomplement local training and on ‑device weight updates. Suchn approach could not only improve energy efficiency butlso offer advantages in terms of privacy and security, andnable user ‑specific personalization of the speech ‑recognitionodule. Finally, multimodal integration—i.e., combining mul-iple task ‑specialized physical reservoir computing units—couldrovide a powerful machine–human interface for analyzing androcessing biosignals from a wide range of advanced sensingevices beyond MEG. In particular, coordinated inference acrossunctional RC units may help mitigate misrecognition causedy posture changes or motion ‑induced variations in throatibrations.  Conclusion n this study, we demonstrated a real-time spoken-digit recog-ition system that integrates a throat-mounted π–gel- electretEG sensor with an ion-gel/graphene-based IGR. The MEGaptures mechanical vibrations generated during articulation,nd the IGR directly processes these raw vibration signals withoutequiring any frequency-domain preprocessing such as MFCCsr cochleagrams. By exploiting the intrinsic nonlinear dynamics,igh-dimensional state expansion, and memory inherent to theGR, the system achieves effective linear separability of speechatterns within a compact physical reservoir. We first character-zed the device properties of the multi-channel graphene IGR andhowed that variations in channel geometry and drain bias yieldiverse nonlinear transfer characteristics suitable for reservoiromputing. When driven by MEG-derived pulse inputs, the IGRroduced eight-dimensional current responses whose temporaldvanced Electronic Materials, 2026evolution encoded digit-dependent features with rich nonlin-ear structure. These reservoir states enabled high classificationperformance even with minimal node counts. Using only 32nodes, the system achieved 84.2% accuracy—substantially out-performing the MEG-only baseline—demonstrating the strongcomputational capacity of the IGR. By increasing the nodesampling resolution, the accuracy further improved to 96.8% with576 virtual nodes. The results highlight several key advantagesof MEG–IGR physical reservoir computing: (i) elimination ofpreprocessing pipelines and spectral analysis, (ii) tunable trade-off between computational performance and reservoir size foredge deployment, (iii) robustness of throat-vibration sensing inchallenging acoustic environments, and (iv) the potential forsecure user authentication based on biomechanical signaturesthat are inherently difficult to spoof. These findings underscorethe promise of IGR-based physical reservoirs as efficient and flexi-ble computational front-ends for next-generation edge-AI sensingdevices. Future work will explore integration of the MEG–IGR platform with on-device learning, multimodal sensing, andbroader classes of vocal and non-vocal human–machine inter-faces, further expanding its applicability to wearable, assistive,and secure edge-intelligent systems. 4 Experimental Section/Methods 4.1 Fabrication of the Gel-Electret MEG Sensor and Recording of Throat-Vibration Signals The mechanoelectric generator-electret sensor utilized in thisstudy was fabricated in a manner similar to a previously reportedmethod [ 55 ]. We keep it brief here. An alkylated-pyrene liquid[ 55 ] was mixed with 1 wt.% of a gelator compound, N , N’ -((1 S ,2 S )-cyclohexane-1,2-diyl)didodecanamide [ 68 ], and soakedinto polyurethane nonwoven fabric (MELTBLOWN NONWO-VENS UC0040, Kuraray Kuraflex Co., Ltd., Japan). Corona-discharging parameters included an applied voltage of + 80 kV,temperature of 100◦C, heating for 30 min, and poling for 90 min.Positive discharging was performed with a corona discharginggun (GC90PN, Green Techno Co., Ltd., Japan). The sensorswere fabricated on polyurethane (Breath Roll BR15, Nitto Co.,Ltd., Japan, 13 µm thick) by applying silver-plated nylon fibers(AGposs, Mitsufuji Co., Ltd., Japan, fiber length = 3 mm), witha piece of copper tape attached at the end to connect to the wire(Figure 1a ) [ 55 ]. Breath Roll was also used as the polyurethanemasking layer. The flexible sensor was fixed just below the Adam’sapple using Scotch mending tape (3 M Company, USA). The open-circuit voltage ( Voc ) of the sensor during speech was recorded onan oscilloscope (MT-775, MotherTool Co., Ltd., Japan) equippedwith a 1 M Ω input impedance probe. The bottom electrode ofthe sensor was connected to the probe, and the top electrode wasgrounded. Speech consisted of uttering each word from zero tonine 100 times, and recording the output on an oscilloscope eachtime. 4.2 Fabrication of the Ion-Gel/Graphene IGR and Electrical Measurements Monolayer graphene grown by chemical vapor deposition (CVD)and transferred onto a SiO2 /Si substrate (300 nm SiO2 ) was9 of 12tive Commons Licensep  t  wc  a  t  e  s  m  T  l  T  c  1  3  1  (  c  w  t  m  c  G  u  (  d  c4A  a  x  u  t  o  s  s  r  n  e u  a  m  t w  t  5  s  a  s  t  i  r  b  e                                       1 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creaurchased from Graphenea (Spain). The graphene was pat-erned by dry etching to form six independent channels:idths of 30 µm with lengths of 5, 20, and 100 µm (ch1–h3), and widths of 80 µm with the same lengths (ch4–ch6),s illustrated in the inset of Figure 2a . Source and drain elec-rodes were defined by photolithography and electron-beamvaporation of Cr/Au (10/50 nm). The ion-gel electrolyte wasynthesized by chemically gelling the ionic liquid 1-ethyl-3-ethylimidazolium bis(trifluoromethylsulfonyl)imide (EMIm-FSI). A polymer solution of poly(dimethylaminoethyl methacry-ate) (PDMEMA) in toluene (200 µL) was mixed with EMIm-FSI (1 mL) under stirring, followed by the addition of theross-linker N,N,N,N’ -tetra(trifluoromethanesulfonyl)-dodecane-,12-diamine (C12 TFSA) (200 µL). After mixing for 30 min at00 rpm, the solution was cast onto a Si wafer and heated at00◦C for 15 min to induce gelation. The resulting ion-gel sheet ∼ 500 µm thick) was cut to size, laminated onto the graphenehannels, and dried overnight in a vacuum chamber evacuatedith a turbomolecular pump. A 1- µm-thick Au foil was placed onop of the gel and used as the common gate electrode. Electricaleasurements were performed at room temperature in a vacuumryostat, with all device contacts made via Al wire bonding.ate voltage pulses were applied using the pulse measurementnit (PMU) of a 4200A-SCS semiconductor parameter analyzerKeithley, USA), enabling time-resolved acquisition of the gate,rain, and source current responses from all six graphenehannels. .3 Readout Weight Training s illustrated in Figure 4a , the temporal output of the reservoir for given time window is obtained from the reservoir state vector according to Equation ( 1 ). Here we describe the proceduresed to compute the readout weight matrix W . Let x1 denotehe reservoir state vector obtained from the first time windowf the first training sample, and x2 , x3 , . . . the vectors obtainedequentially from all subsequent time windows of all trainingamples. By stacking these vectors in order, we construct theeservoir state matrix X = ( x1 , x2 ,. . . , xM )T where 𝑀 =∑400 𝑖= 1 𝑛𝑖 andi is the number of time windows for the i -th training sample. Forach state vector xj , we define a corresponding teacher vector yt, jsing a one-hot representation of the ground-truth digit label. Byrranging these teacher vectors row-wise, we obtain the teacheratrix Yt = ( yt,1 , yt,2 ,. . . , yt, M )T The readout weight matrix W wasrained using ridge regression, 𝑊 = 𝑌t 𝑋T (𝑋𝑋T + 𝜆𝐼 )− 1 (2)here λ = 0.001 is the regularization parameter and 𝐼 ∈ ℝ𝑀×𝑀 ishe identity matrix. Classification accuracy was evaluated using-fold cross-validation at the utterance (sample) level, with 400amples used for training and 100 for testing in each split. Becausell 500 utterances were acquired from a single speaker in aingle recording session under identical sampling conditions,he train/test partitioning is not confounded by inter-speaker ornter-session variability. Nevertheless, to further strengthen theobustness of the evaluation and to eliminate any unintendedias, the utterances were randomly partitioned into five mutuallyxclusive folds while maintaining class balance (digits 0–9) in0 of 12each split. Importantly, the partitioning was performed withoutoverlap so that (i) no utterance appeared in both the training andtest sets within a fold, and (ii) each utterance was used as test dataexactly once across the five folds. Acknowledgements This research was in part supported by JST PRESTO Grant numberJPMJPR23H4 and JSPS KAKENHI Grant Number JP24KJ0299 (Grant-in-Aid for JSPS Fellows), JP25K17941, and JP25H01264 (Grant-in-Aid forTransformative Research Areas “π-Molecular Complexity”). A part of thiswork was supported by “Advanced Research Infrastructure for Materialsand Nanotechnology in Japan (ARIM)” of the Ministry of Education,Culture, Sports, Science and Technology (MEXT). This work was partiallysupported by Innovative Science and Technology Initiative for SecurityGrant Number JPJ004596, ATLA, Japan. We also acknowledge theWorld Premier International Research Center Initiative (WPI), MEXT,Japan. Proposal Number JPMXP1224NM5236 and JPMXP1225NM5329.D.N. acknowledges the International Center for Young Scientists (ICYS)at NIMS for providing ICYS fellowships. We thank Mr. Kota Fukuda(NIMS) for his assistance in acquiring data from the π–gel- electret MEGsensor. Funding This work was supported by JST PRESTO Grant number JPMJPR23H4,JSPS KAKENHI Grant Number JP24KJ0299 (Grant-in-Aid for JSPSFellows), JP25K17941, JP25H01264 (Grant-in-Aid for TransformativeResearch Areas “π-Molecular Complexity”), and Innovative Science andTechnology Initiative for Security Grant Number JPJ004596, ATLA,Japan. Conflicts of Interest The authors declare no conflict of interest. Data Availability Statement The data that support the findings of this study are available from thecorresponding author upon reasonable request. References 1 . J. M. Baker, L. Deng, J. Glass, et al., “Developments and Directions inSpeech Recognition and Understanding, Part 1 [DSP Education],” IEEESignal Processing Magazine 26 (2009): 75–80, https://doi.org/10.1109/MSP.2009.932166 . 2 . S. Davis and P. Mermelstein, “Comparison of Parametric Represen-tations for Monosyllabic Word Recognition in Continuously SpokenSentences,” IEEE Transactions on Acoustics, Speech, and Signal Processing28 (1980): 357–366, https://doi.org/10.1109/TASSP.1980.1163420 . 3 . W. Han, C.-F. Chan, C.-S. Choy, and K.-P. Pun, “An Efficient MFCCExtraction Method in Speech Recognition,” in 2006 IEEE InternationalSymposium on Circuits and Systems (ISCAS) (IEEE, 2006). 4 . G. Hinton, L. Deng, D. Yu, et al., “Deep Neural Networks for AcousticModeling in Speech Recognition: The Shared Views of Four ResearchGroups,” IEEE Signal Processing Magazine 29 (2012): 82–97, https://doi.org/10.1109/MSP.2012.2205597 . 5 . Y. Gong, “Speech Recognition in Noisy Environments: A Survey,”Speech Communication 16 (1995): 261–291, https://doi.org/10.1016/0167- 6393(94)00059-J . 6 . Q. Li, Y. Yang, T. Lan, et al., “MSP-MFCC: Energy-Efficient MFCCFeature Extraction Method with Mixed-Signal Processing Architecturefor Wearable Speech Recognition Applications,” IEEE Access 8 (2020):4 8720–4 8730, https://doi.org/10.1109/ACCESS.2020.2979799 . Advanced Electronic Materials, 2026tive Commons Licensehttps://doi.org/10.1109/MSP.2009.932166https://doi.org/10.1109/TASSP.1980.1163420https://doi.org/10.1109/MSP.2012.2205597https://doi.org/10.1016/0167-6393(94)00059-Jhttps://doi.org/10.1109/ACCESS.2020.29797997  t  h8  L  v  j9  R  C11  t  E1  M  n1  a  C51  S  h1  AA  21  S  &  1  B  S  (1  UA  a1  CO  21  S  (2  P  12  P  C2  C  s2  L  W  d2  P  v  1                                             A 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Cre . Y. Patel, S. Tanwar, R. Gupta, et al., “Deepfake Generation and Detec-ion: Case Study and Challenges,” IEEE Access 11 (2023): 143296–143323,ttps://doi.org/10.1109/ACCESS.2023.3342107 .  . Z. Wu, N. Evans, T. Kinnunen, J. Yamagishi, F. Alegre, and H.i, “Spoofing and Countermeasures for Speaker Verification: A Sur-ey,” Speech Communication 66 (2015): 130–153, https://doi.org/10.1016/.specom.2014.10.005 .  . S. Lee, J. Kim, I. Yun, et al., “An Ultrathin Conformable Vibration-esponsive Electronic Skin for Quantitative Vocal Recognition,” Natureommunications 10 (2019): 2468, https://doi.org/10.1038/s41467-019- 0465-w. 0 . S. Gao, G. Li, S. Wu, et al., “High-Bandwidth Self-Powered Vibra-ion Sensors Based on Triboelectric Particle-Surface Interactions,” Nanonergy 119 (2024): 109060, https://doi.org/10.1016/j.nanoen.2023.109060 . 1 . Q. Zhang and K. Tong, “Vocal Cord Vibration Signal Recognitionodel Based on Feature Engineering Preprocessing,” IEEE Sensors Jour-al 23 (2023): 31380–31388, https://doi.org/10.1109/JSEN.2023.3321987 . 2 . T. Liu, M. Zhang, Z. Li, et al., “Machine Learning-Assisted Wear-ble Sensing Systems for Speech Recognition and Interaction,” Natureommunications 16 (2025): 2363, https://doi.org/10.1038/s41467-025- 7629-5 . 3 . J. Zhao, D. Chen, L. Zhao, et al., “Self-Powered Speech Recognitionystem for Deaf Users,” Cell Reports Physical Science 3 (2022): 101168,ttps://doi.org/10.1016/j.xcrp.2022.101168 . 4 . S. Lee, H. Roh, J. Kim, et al., “An Electret-Powered Skin-Attachableuditory Sensor That Functions in Harsh Acoustic Environments,”dvanced Materials 34 (2022): 2205537, https://doi.org/10.1002/adma.02205537 . 5 . W. Li, R. Min, D. Zheng, et al., “Wearable Photonic Artificial Throat forilent Communication and Speech Recognition,” ACS Applied Materials Interfaces 17 (2025): 11126–11143, https://doi.org/10.1021/acsami.4c21754 .6 . K. Xiao, S. Zhou, H. Wang, X. Li, R. Min, and Z. Wang, “Photoniciointerfaces: A New Paradigm for Seamless Integration with Livingystems Based on Flexible Optical Fibers,” Optics & Laser Technology 1972026): 114776, https://doi.org/10.1016/j.optlastec.2026.114776 . 7 . R. Zhong, J. Pan, Q. Li, et al., “Cuffless Wearable Sphy gmomanometerssing Dual Digital Optical Frequency Combs in Chalcogenide Chips,”dvanced Functional Materials 35 (2025): 2417142, https://doi.org/10.1002/dfm.202417142 . 8 . W. Li, Y. Long, Y. Yan, et al., “Wearable Photonic Smart Wristband forardiorespiratory Function Assessment and Biometric Identification,”pto-Electronic Advances 8 (2025): 240254, https://doi.org/10.29026/oea.025.240254 . 9 . H. Jaeger and H. Haas, “Harnessing Nonlinearity: Predicting Chaoticystems and Saving Energy in Wireless Communication,” Science 3042004): 78–80, https://doi.org/10.1126/science.1091277 . 0 . G. Tanaka, T. Yamane, J. B. Héroux, et al., “Recent Advances inhysical Reservoir Computing: A Review,” Neural Networks 115 (2019):00–123, https://doi.org/10.1016/J.NEUNET.2019.03.005 . 1 . L. Appeltant, M. C. Soriano, G. Van der Sande, et al., “Informationrocessing Using a Single Dynamical Node as Complex System,” Natureommunications 2 (2011): 468, https://doi.org/10.1038/ncomms1476 . 2 . Y. Paquot, F. Duport, A. Smerieri, et al., “Optoelectr onic Reservoiromputing,” Scientific Reports 2 (2012): 287, https://doi.org/10.1038/rep00287 . 3 . R. Martinenghi, S. Rybalko, M. Jacquot, Y. K. Chembo, and L.arger, “Photonic Nonlinear Transient Computing with Multiple-Delayavelength Dynamics,”Physical Review Letters 108 (2012): 244101, https://oi.org/10.1103/PhysRevLett.108.244101 . 4 . L. Larger, M. C. Soriano, D. Brunner, et al., “Photonic Informationrocessing beyond Turing: An Optoelectronic Implementation of Reser-oir Computing,” Optics Express 20 (2012): 3241–3249, https://doi.org/10.364/OE.20.003241 . dvanced Electronic Materials, 202625 . L. Larger, A. Baylón-Fuentes, R. Martinenghi, V. S. Udaltsov, Y. K.Chembo, and M. Jacquot, “High-Speed Photonic Reservoir ComputingUsing a Time-Delay-Based Architecture: Million Words per SecondClassification,” Physical Review X 7 (2017): 11015, https://doi.org/10.1103/PhysRevX.7.011015 . 26 . J. Torrejon, M. Riou, F. A. Araujo, et al., “Neuromorphic Computingwith Nanoscale Spintronic Oscillators,” Nature 547 (2017): 428–431,https://doi.org/10.1038/nature23011 . 27 . Y. Zhong, J. Tang, X. Li, B. Gao, H. Qian, and H. Wu, “DynamicMemristor-Based Reservoir Computing for High-Efficiency TemporalSignal Processing,”Nature Communicators 12 (2021): 408, https://doi.org/10.1038/s41467- 020- 20692- 1 . 28 . S. Lilak, W. Woods, K. Scharnhorst, et al., “Spoken Digit Classificationby In-Materio Reservoir Computing with Neuromorphic Atomic SwitchNetworks,” Frontiers in Nanotechnology 3 (2021): 675792, https://doi.org/10.3389/fnano.2021.675792 . 29 . J. Moon, W. Ma, J. H. Shin, et al., “Temporal Data Classificationand Forecasting Using a Memristor-Based Reservoir Computing System,”Nature Electronics 2 (2019): 4 80–4 87, https://doi.org/10.1038/s41928-019- 0313-3 . 30 . R. Midya, Z. Wang, S. Asapu, et al., “Reservoir Computing UsingDiffusive Memristors,” Advanced Intelligent Systems 1 (2019): 1900084,https://doi.org/10.1002/aisy.201900084 . 31 . C. Du, F. Cai, M. A. Zidan, W. Ma, S. H. Lee, and W. D. Lu, “ReservoirComputing Using Dynamic Memristors for Temporal Information Pro-cessing,” Nature Communicators 8 (2017): 2204, https://doi.org/10.1038/s41467- 017- 02337- y. 32 . Y. Usami, B. van de Ven, D. G. Mathew, et al., “In-MaterioReservoir Computing in a Sulfonated Polyaniline Network,”Advanced Materials 33 (2021): 2102688, https://doi.org/10.1002/adma.202102688 . 33 . G. Milano, G. Pedretti, K. Montano, et al., “In Materia Reservoir Com-puting with a Fully Memristive Architecture Based on Self-OrganizingNanowire Networks,” Nature Materials 21 (2022): 195–202, https://doi.org/10.1038/s41563- 021- 01099- 9 . 34 . O. Lee, T. Wei, K. D. Stenning, et al., “Task-Adaptive Physical ReservoirComputing,” Nature Materials 23 (2024): 79–87, https://doi.org/10.1038/s41563- 023- 01698- 8 . 35 . J. Hochstetter, R. Zhu, A. Loeffler, A. Diaz-Alvarez, T. Nakayama, andZ. Kuncic, “Avalanches and Edge-of-Chaos Learning in NeuromorphicNanowire Networks,” Nature Communications 12 (2021): 4008, https://doi.org/10.1038/s41467- 021- 24260- z . 36 . S. Kan, K. Nakajima, T. Asai, and M. Akai-Kasaya, “PhysicalImplementation of Reservoir Computing through Electrochemical Reac-tion,” Advanced Science 9 (2022): 2104076, https://doi.org/10.1002/advs.202104076 . 37 . K. Toprasertpong, E. Nako, Z. Wang, R. Nakane, M. Takenaka, and S.Takagi, “Reservoir Computing on a Silicon Platform with a FerroelectricField-Effect Transistor,”Communications Engineering 1 (2022): 21, https://doi.org/10.1038/s44172-022-00021-8 . 38 . W. Namiki, Y. Yamaguchi, D. Nishioka, T. Tsuchiya, and K.Terabe, “Opto-Magnonic Reservoir Computing Coupling NonlinearInterfered Spin Wave and Visible Light Switching,” MaterialsToday Physics 45 (2024): 101465, https://doi.org/10.1016/j.mtphys.2024.101465 . 39 . W. Namiki, D. Nishioka, Y. Yamaguchi, T. Tsuchiya, T. Higuchi, andK. Terabe, “Experimental Demonstration of High-Performance PhysicalReservoir Computing with Nonlinear Interfered Spin Wave Multidetec-tion,” Advanced Intelligent Systems 5 (2023): 2300228, https://doi.org/10.1002/aisy.202300228 . 40 . W. Namiki, D. Nishioka, Y. Nomura, T. Tsuchiya, K. Yamamoto, and K.Terabe, “Iono–Magnonic Reservoir Computing with Chaotic Spin WaveInterference Manipulated by Ion-Gating,” Advanced Science 12 (2024):2411777, https://doi.org/10.1002/advs.202411777 . 11 of 12ative Commons Licensehttps://doi.org/10.1109/ACCESS.2023.3342107https://doi.org/10.1016/j.specom.2014.10.005https://doi.org/10.1038/s41467-019-10465-whttps://doi.org/10.1016/j.nanoen.2023.109060https://doi.org/10.1109/JSEN.2023.3321987https://doi.org/10.1038/s41467-025-57629-5https://doi.org/10.1016/j.xcrp.2022.101168https://doi.org/10.1002/adma.202205537https://doi.org/10.1021/acsami.4c21754https://doi.org/10.1016/j.optlastec.2026.114776https://doi.org/10.1002/adfm.202417142https://doi.org/10.29026/oea.2025.240254https://doi.org/10.1126/science.1091277https://doi.org/10.1016/J.NEUNET.2019.03.005https://doi.org/10.1038/ncomms1476https://doi.org/10.1038/srep00287https://doi.org/10.1103/PhysRevLett.108.244101https://doi.org/10.1364/OE.20.003241https://doi.org/10.1103/PhysRevX.7.011015https://doi.org/10.1038/nature23011https://doi.org/10.1038/s41467-020-20692-1https://doi.org/10.3389/fnano.2021.675792https://doi.org/10.1038/s41928-019-0313-3https://doi.org/10.1002/aisy.201900084https://doi.org/10.1038/s41467-017-02337-yhttps://doi.org/10.1002/adma.202102688https://doi.org/10.1038/s41563-021-01099-9https://doi.org/10.1038/s41563-023-01698-8https://doi.org/10.1038/s41467-021-24260-zhttps://doi.org/10.1002/advs.202104076https://doi.org/10.1038/s44172-022-00021-8https://doi.org/10.1016/j.mtphys.2024.101465https://doi.org/10.1002/aisy.202300228https://doi.org/10.1002/advs.2024117774  A  v  a4  T  S  S4  T  L  24  K  E  S  h4  T  D  34  o  Y  N  h4  fA  04  “  M  a4  “  I  S  s5  S  AN  h5  C  a  E5  v  D  (5  T  A  N5  “  C  A  a5F            T                         1 2199160x, 2026, 19, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/aelm.202600006 by Takashi Tsuchiya - National Institute For , Wiley Online Library on [14/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applica1 . D. Nishioka, T. Tsuchiya, W. Namiki, et al., “Edge-of-Chaos Learningchieved by Ion-Electron–Coupled Dynamics in an Ion-Gating Reser-oir,” Science Advances 8 (2022): ade1156, https://doi.org/10.1126/sciadv.de1156 . 2 . T. Wada, D. Nishioka, W. Namiki, T. Tsuchiya, T. Higuchi, and K.erabe, “A Redox-Based Ion-Gating Reservoir, Utilizing Double Reservoirtates in Drain and Gate Nonlinear Responses,” Advanced Intelligentystems 5 (2023): 2300123, https://doi.org/10.1002/aisy.202300123 . 3 . K. Shibata, D. Nishioka, W. Namiki, T. Tsuchiya, T. Higuchi, and K.erabe, “Redox-Based Ion-gating Reservoir Consisting of (104) OrientediCoO2 Film, Assisted by Physical Masking,” Scientific Reports 13 (2023):1060, https://doi.org/10.1038/s41598- 023- 48135-z . 4 . D. Nishioka, Y. Shingaya, T. Tsuchiya, T. Higuchi, and. Terabe, “Few- and Single-Molecule Reservoir Computingxperimentally Demonstrated with Surface-Enhanced Ramancattering and Ion Gating,” Science Advances 10 (2024): adk6438,ttps://doi.org/10.1126/sciadv.adk6438 . 5 . D. Nishioka, T. Tsuchiya, M. Imura, Y. Koide, T. Higuchi, and K.erabe, “A High-performance Deep Reservoir Computer Experimentallyemonstrated with Ion-Gating Reservoirs,”Communications Engineering (2024): 81, https://doi.org/10.1038/s44172-024-00227-y. 6 . M. Takayanagi, D. Nishioka, T. Tsuchiya, et al., “Ultrafast-Switchingf an All-Solid-State Electric Double Layer Transistor with a Porousttria-Stabilized Zirconia Proton Conductor and the Application toeuromorphic Computing,”Materials Today Advances 18 (2023): 100393,ttps://doi.org/10.1016/j.mtadv.2023.100393 . 7 . Y. Yamaguchi, D. Nishioka, W. Namiki, et al., “Inverted Input Methodor Computing Performance Enhancement of the Ion-Gating Reservoir,”pplied Physics Express 17 (2024): 024501, https://doi.org/10.35848/1882-786/ad2906 . 8 . W. Namiki, D. Nishioka, T. Tsuchiya, T. Higuchi, and K. Terabe,Magnetization Vector Rotation Reservoir Computing Operated by Redoxechanism,” Nano Letters 24 (2024): 4383–4392, https://doi.org/10.1021/cs.nanolett.3c05029 . 9 . T. Tsuchiya, M. Takayanagi, D. Nishioka, W. Namiki, and K. Terabe,Electric Double Layer Effect in the Vicinity of Solid Electrolyte/Diamondnterfaces and the Application to Neuromorphic Computing,” Journal ofolid State Electrochemistry 28 (2024): 4367–4376, https://doi.org/10.1007/10008- 024- 05937- z . 0 . D. Nishioka, T. Tsuchiya, T. Higuchi, and K. Terabe, “Enhancedynaptic Characteristics of Hx WO3 -Based Neuromorphic Devices,chieved by Current Pulse Control, for Artificial Neural Networks,”euromorphic Computing and Engineering 3 (2023): 034008,ttps://doi.org/10.1088/2634-4386/acf1c6 . 1 . H. Kitano, D. Nishioka, K. Terabe, and T. Tsuchiya, “Physical Reservoiromputing with Graphene-Based Solid Electric Double Layer Transistornd the Information Processing Capacity Analysis,” Applied Physicsxpress 18 (2025): 024501, https://doi.org/10.35848/1882-0786/adb19b . 2 . T. Tsuchiya, D. Nishioka, W. Namiki, and K. Terabe, “Physical Reser-oir Computing Utilizing Ion-Gating Transistors Operating in Electricouble Layer and Redox Mechanisms,” Advanced Electronic Materials 102024): 2400625, https://doi.org/10.1002/aelm.202400625 . 3 . D. Nishioka, H. Kitano, W. Namiki, S. Souma, K. Terabe, and T.suchiya, “Two Orders of Magnitude Reduction in Computational Loadchieved by Ultrawideband Responses of an Ion-Gating Reservoir,” ACSano 19 (2025): 36896–36914, https://doi.org/10.1021/acsnano.5c06174 . 4 . D. Nishioka, K. Shibata, W. Namiki, K. Terabe, and T. Tsuchiya,Physical Masking-Induced Enhancement of Information Processingapacity in a Redox-Type Ion-Gating Reservoir,” Japanese Journal ofpplied Physics 64 (2025): 11SP25, https://doi.org/10.35848/1347-4065/e1d84 . 5 . A. Tateyama, K. Nagura, M. Yamanaka, and T. Nakanishi, “Alkyl–πunctional Molecular Gels: Control of Elastic Modulus and Improvement2 of 12of Electret Performance,” Angewandte Chemie International Edition 63(2024): 202402874, https://doi.org/10.1002/anie.202402874 . 56 . A. Ghosh, M. Yoshida, K. Suemori, et al., “Soft Chromophore FeaturedLiquid Porphyrins and Their Utilization toward Liquid Electret Applica-tions,” Nature Communications 10 (2019): 4210, https://doi.org/10.1038/s41467- 019- 12249-8 . 57 . R. K. Gupta, M. Yoshida, A. Saeki, Z. Guo, and T. Nakanishi,“Alkyl-C60 Liquid Electrets as Deformable Mechanoelectric Genera-tors,” Materials Horizons 10 (2023): 3458–3466, https://doi.org/10.1039/D3MH00485F . 58 . A. Shinohara, M. Yoshida, C. Pan, and T. Nakanishi, “Stretchable π-Conjugated Polymer Electrets for Mechanoelectric Generators,” PolymerJournal 55 (2023): 529–535, https://doi.org/10.1038/s41428-022- 00725- w. 59 . D. Nishioka, T. Tsuchiya, T. Higuchi, and K. Terabe, “Oxygen-olerant Operation of All-Solid-State Ionic-Gating Devices: Advantage ofAll-Solid-State Structure for Ionic-Gating,” Japanese Journal of AppliedPhysics 59 (2020): SIIG09, https://doi.org/10.35848/1347-4065/ab7e12 . 60 . M. Takayanagi, T. Tsuchiya, D. Nishioka, et al.,“Accelerated/Decelerated Dynamics of the Electric Double Layerat Hydrogen-Terminated Diamond/Li+ Solid Electrolyte Interface,”Materials Today Physics 31 (2023): 101006, https://doi.org/10.1016/j.mtphys.2023.101006 . 61 . T. Tsuchiya, M. Takayanagi, K. Mitsuishi, et al., “The ElectricDouble Layer Effect and Its Strong Suppression at Li+ Solid Elec-trolyte/Hydrogenated Diamond Interfaces,” Communications Chemistry4 (2021): 117, https://doi.org/10.1038/s42004- 021- 00554-7 . 62 . H. Wang, Y. Wu, C. Cong, J. Shang, and T. Yu, “Hysteresis of ElectronicTransport in Graphene Transistors,” ACS Nano 4 (2010): 7221–7228,https://doi.org/10.1021/nn101950n . 63 . A. Tries, N. Richter, Z. Chen, et al., “Hysteresis in Graphene Nanorib-bon Field-Effect Devices,” Physical Chemistry Chemical Physics 22 (2020):5667–5672, https://doi.org/10.1039/D0CP00298D . 64 . M. C. Lemme, T. J. Echtermeyer, M. Baus, and H. Kurz, “A GrapheneField-Effect Device,” IEEE Electron Device Letters 28 (2007): 282–284,https://doi.org/10.1109/LED.2007.891668 . 65 . S. Zhao, B. Zhang, J. Yang, J. Zhou, and Y. Xu, “Linear DiscriminantAnalysis,” Nature Reviews Methods Primers 4 (2024): 70, https://doi.org/10.1038/s43586-024-00346-y. 66 . J. Dambre, D. Verstraeten, B. Schrauwen, and S. Massar, “InformationProcessing Capacity of Dynamical Systems,” Scientific Reports 2 (2012):514. 67 . M. Zolfagharinejad, J. Büchel, L. Cassola, et al., “Analogue SpeechRecognition Based on Physical Computing,” Nature 645 (2025): 886–892,https://doi.org/10.1038/s41586- 025- 09501- 1 . 68 . K. Hanabusa, M. Yamada, M. Kimura, and H. Shirai, “ProminentGelation and Chiral Aggregation of Alkylamides Derived from Trans -1,2-Diaminocyclohexane,” Angewandte Chemie International Edition inEnglish 35 (1996): 1949–1951, https://doi.org/10.1002/anie.199619491 . Supporting Information Additional supporting information can be found online in the SupportingInformation section. Supporting File : aelm70410-sup-0001-SuppMat.docx. Advanced Electronic Materials, 2026ble Creative Commons Licensehttps://doi.org/10.1126/sciadv.ade1156https://doi.org/10.1002/aisy.202300123https://doi.org/10.1038/s41598-023-48135-zhttps://doi.org/10.1126/sciadv.adk6438https://doi.org/10.1038/s44172-024-00227-yhttps://doi.org/10.1016/j.mtadv.2023.100393https://doi.org/10.35848/1882-0786/ad2906https://doi.org/10.1021/acs.nanolett.3c05029https://doi.org/10.1007/s10008-024-05937-zhttps://doi.org/10.1088/2634-4386/acf1c6https://doi.org/10.35848/1882-0786/adb19bhttps://doi.org/10.1002/aelm.202400625https://doi.org/10.1021/acsnano.5c06174https://doi.org/10.35848/1347-4065/ae1d84https://doi.org/10.1002/anie.202402874https://doi.org/10.1038/s41467-019-12249-8https://doi.org/10.1039/D3MH00485Fhttps://doi.org/10.1038/s41428-022-00725-whttps://doi.org/10.35848/1347-4065/ab7e12https://doi.org/10.1016/j.mtphys.2023.101006https://doi.org/10.1038/s42004-021-00554-7https://doi.org/10.1021/nn101950nhttps://doi.org/10.1039/D0CP00298Dhttps://doi.org/10.1109/LED.2007.891668https://doi.org/10.1038/s43586-024-00346-yhttps://doi.org/10.1038/s41586-025-09501-1https://doi.org/10.1002/anie.199619491 Ion-Gating Reservoir Computing for Preprocessing-Free Speech Recognition from Throat Vibrations 1 | Introduction 2 | Results and Discussion 2.1 | Concept and Task Design for MEG-IGR-Based Speech Recognition 2.2 | Electrical Characteristics of the Graphene-Based IGR and Its Nonlinear Mapping of MEG Outputs 2.3 | Spoken-Digit Recognition Using MEG-IGR Physical Reservoir Computing 2.4 | Outlook 3 | Conclusion 4 | Experimental Section/Methods 4.1 | Fabrication of the Gel-Electret MEG Sensor and Recording of Throat-Vibration Signals 4.2 | Fabrication of the Ion-Gel/Graphene IGR and Electrical Measurements 4.3 | Readout Weight Training Acknowledgements Funding Conflicts of Interest Data Availability Statement References Supporting Information