An in-sensor communication electronic textile for imperceptible and ultrarobust silent speech Authors: Yuchen Lin, Shifan Yu, Lei Liu, Huasen Wang, Yuhui Wang, Mengting Jiang, Gantang Su, Yue Yu, Hongyu Chen, Yanhao Luo, Zijian Huang, Chenwei Li, Junzhi Yu, Ziquan Guo, Yuhan Su, Zhong Chen, Lichao Yang, Zhenyu Zhao, Qingliang Liao, Yuanjin Zheng, Xinqin Liao Published: 2026 ABSTRACT Aphasiacs with impaired vocal organs suffer profound communication barriers, making silent speech decoding essential. Although audible speech is absent, subtle multi-muscle activities are retained. This provides a physiological basis for the sensing modality that reliably decodes silent speech. Yet conventional microphone and piezoelectric films remain susceptible to dispersion–fidelity conflict, external disturbance in the complex environments, and are further constrained by sensing–communication separation. Here, we report an in-sensor communication scarf interface that imperceptibly captures multi-muscle activities and ultrarobustly decodes silent speech. By fusing spoof surface plasmon–driven in-sensor electromagnetic physiological communication with metamaterial topological embroidery, the system establishes exquisite reconciliation between dispersion-regulated spatial selectivity and ultrahigh fidelity (7 fs²), enabling phase-resolved differentiation and precise capture of multi-muscle activities. This minimalist yet informative rich component design radically transcends limitations of sensing–communication separation, microphone, and piezoelectric films, and greatly delivers anti-motion artifact by 24,300%, anti-electromagnetic interference by 46.4 dB, and anti-noise by 33.44 dB in the complex environments. Combining with machine learning, the interface achieves 98.7% recognition and supports assistive interaction and sleep apnea monitoring. The interface bridges metamaterial electromagnetics and human electrophysiology, establishing a transformative in-sensor physiological decoding paradigm for electromagnetic physiological communication and adaptive myogenic interfacing. Introduction Aphasiacs with impaired vocal organs suffer profound communication barriers. The absence of vocalization not only impedes communication but also triggers psychological and social complications such as depression and isolation1. At a population level, tens of millions of people worldwide are affected, primarily after stroke or traumatic injury, resulting in long-term communication barriers2,3. More seriously, the need for communication also forces them into long-term dependence on clinical support and health monitoring4–6. This long-term clinical reliance disrupts daily routines and imposes substantial burdens on patients and caregivers. Although aphasiacs cannot produce audible speech, they often retain subtle and residual multi-muscle activities when attempting to vocalize7,8. These subtle multi-muscle activities preserve silent speech intent and offer a physiological basis for silent speech decoding. Therefore, decoding of silent speech through subtle multi-muscle activities is essential for both daily interaction and clinical assistance of aphasiacs. However, capturing such subtle multi-muscle activities and decoding for silent speech in real-world complex environments remains a key challenge. Silent speech decoding requires both spatial selectivity to separate multiple muscles and phase integrity to maintain high signal fidelity. Yet existing speech-sensing modalities cannot simultaneously satisfy these two requirements. Existing sensing modalities for speech decoding, primarily microphone and piezoelectric films, are fundamentally misaligned with the requirements of silent speech decoding. As the relevant muscles of silent speech generate subtle and localized contraction patterns in the neck region, requiring dispersion-regulated spatial selectivity to distinguish subtle multi-muscle activities9,10. Spatial selectivity refers to a sensor’s ability to separate signals generated by different muscles within the same region11,12. However, dispersion unavoidably perturbs phase characterization, impairing the phase integrity required for high-fidelity speech processing13–15. This creates an intrinsic conflict in silent speech decoding: increased dispersion improves spatial selectivity, whereas high fidelity needed for phase integrity demands decreased dispersion. In addition to the inability to resolve intrinsic conflict, the more serious problem is that existing sensing modalities for speech decoding suffer severe distortions due to the limitations arising from their transduction mechanisms and external disturbances. Microphones, which operate through diaphragm-based mechano-electrical transduction, are intrinsically dependent on airborne propagation, rendering their signals highly susceptible to severe distortion under noise and electromagnetic interference (EMI)16–19. By contrast, piezoelectric films become the widely-adopted solution for wearable sensing owing to their flexibility and sensitivity20–23. However, their direct mechano-electrical transduction is critically dependent on intimate skin-device contact. But pores, perspiration, and continuous muscle motion destabilize the interface and cause severe distortion. Except for these sensing modalities, existing systems separate sensing and communication into isolated modules, producing discrete structures with redundant components and deficient data that are susceptible to interference and prone to distortion24–26. This separation further disrupts the phase integrity required for decoding. As a result, reliable silent speech decoding still requires a co-designed sensing–communication textile, enabling high fidelity, moderate dispersion, and comfortable long-term wearability27–29. Yet no existing system achieves it. In this article, we report an in-sensor communication (ISC) scarf interface that imperceptibly captures the subtle multi-muscle activities and ultrarobustly decodes silent speech generated when aphasiacs attempt to vocalize. The textile fusing spoof surface plasmon (SSP)–driven in-sensor electromagnetic physiological communication with metamaterial topological embroidery design. These two achieve practical optimization of the tradeoff between dispersion-regulated spatial selectivity and ultrahigh fidelity (7 fs²), enabling precise signals of multi-muscle activities. We find that spatial information differences arising from the activity of distinct muscle groups are primarily manifested in the phase of the propagating surface wave. Benefiting from the SSP-enabled confined-field propagation and embroidery-governed modal shaping, the interface stably tracks distinct muscle–phase signatures even when activities amplitude is tiny or partially overlapped. Radically transcending limitations of sensing–communication separation, the ISC provides immunity to interference of transmission, but also a minimalist component with rich informative data. This ISC framework avoids data losses and impedance mismatches across interfaces, allowing physiological waveforms to be preserved and decoded without intermediate conversion, buffering, or analog deterioration. Transcending the inherent physical limitations of acoustic and piezoelectric, the ISC scarf interface greatly delivers anti-motion artifact by 24,300%, and anti-EMI by 46.4 dB. It also maintains anti-noise with a signal–to–noise ratio (SNR) of 33.44 dB in complex environments. Even in purely random human activities and complex environments, the metamaterial-guided conduction pathway enables stable reading of subtle myogenic signals, reflecting ultrarobust adaptability to real-world usage rather than laboratory isolation. Combined with an efficient machine learning model, the system achieves 98.7% recognition in precise silent speech decoding. The imperceptible textile architecture further supports long-term wear, breathability, and steady temperature. Capturing subtle multi-muscle activities, this interface further validates its reliability in physiological applications like sleep apnea monitoring and assistive interaction. By bridging metamaterial electromagnetics with human electrophysiology, this work establishes a transformative in-sensor physiological decoding paradigm for electromagnetic physiological communication and adaptive myogenic interfacing. Results Conceptual framework and functional advantages of the ISC scarf interface Aphasiacs retained the function of articulatory muscles, which include the neck groups, yet were unable to generate acoustic output due to impaired vocal organs, leading to severe barriers in daily communication and psychological well-being. These preserved multi-muscle activities provided a physiological basis for silent speech decoding, motivating us to develop a mechanism that directly converts these subtle activities into phase characterization and then decoding. To achieve this, the ISC scarf interface generated a SSP mode along a metamaterial topological embroidery pathway (Fig. 1a and Supplementary Fig. 1). When excited, the SSP supports a transverse propagation and longitudinal attenuation with a tightly confined field30,31. This near-field component interacts with the superficial articulatory muscles that underlie silent speech. During attempted vocalization, the contraction and approaches of muscles, such as the masseter, suprahyoid muscle complex, and thyrohyoid modulate the electromagnetic environments (Fig. 1b). At the microscopic level, the incident SSP field induces a reverse-polarization response within the muscle tissue due to its high-water content and anisotropic permittivity. This produces power absorption through lossy dielectric coupling, leading to further attenuation of the confined field and slight redistribution of local energy density. At the macroscopic level, this manifests as a measurable decay in the transmitted signal amplitude and phase change that reflects the instantaneous mechanical state of the muscles. These processes correspond to the progression of (1) signal input, (2) power absorption, (3) signal attenuation, and (4) signal output. The transmission pattern with and without the human body further validates these propagation characteristics (Fig. 1c, d). Without a human body, the SSP propagates along the textile, maintaining a symmetric confined field profile. With the presence of the human body, however, the field concentrated near the tissue–interface region due to confined field coupling, producing a clear contrast in the propagation pathway. This interface redistributed the confined field energy density and modifies the phase of the surface-wave mode. Consequently, variations in multi-muscle activities or physiological state were encoded as measurable changes in the detected signal, establishing a direct electromagnetic pathway between muscle activity and signal modulation. This mechanism formed the physical basis of muscle–wave coupling and enabled high-sensitivity detection of subtle multi-muscle activities without requiring direct mechanical contact. The phase variations generated through confined field, imperceptible interaction, were processed and input into a signal-processing module and a convolutional neural network to decode the silent speech. Existing microphones and piezoelectric films faced numerous limitations in addressing silent speech decoding, which motivated us to explore alternative mechanisms and to perform comparative studies (Fig. 1e–h).Fig. 1Concept of the ISC scarf interface and comparative advantages.a Illustration of the ISC scarf interface for an aphasiac. Wireless signals interact imperceptibly with the neck of the aphasiac through the ISC scarf interface for silent speech decoding (ISC in-sensor communication) (Contains photos by Juicy_fish via Magnific License). b Conceptual image of the ISC scarf interface interacting with the laryngeal mandibular muscles (Repeat 3 times with similar results). c Electric field distribution during normal propagation and electric absorption (side view) (E electric field) (The value of |E|max is 1.18 × 104 V·m−1). d Magnetic field distribution during normal propagation and magnetic absorption (side view) (H magnetic field) (The value of |H|max is 3.11 × 101 A·m−1). e Contrast of motion artifact between existing modalities and the ISC scarf interface. f Contrast of noise between existing modalities and the ISC scarf interface. g Contrast of EMI between existing modalities and the ISC scarf interface (EMI electromagnetic interference) (Contains photos by Freepik via Magnific License). h Contrast of comfort between existing modalities and the ISC scarf interface. (Contains photos by studiogstock via Magnific License). Concept of the ISC scarf interface and comparative advantages. a Illustration of the ISC scarf interface for an aphasiac. Wireless signals interact imperceptibly with the neck of the aphasiac through the ISC scarf interface for silent speech decoding (ISC in-sensor communication) (Contains photos by Juicy_fish via Magnific License). b Conceptual image of the ISC scarf interface interacting with the laryngeal mandibular muscles (Repeat 3 times with similar results). c Electric field distribution during normal propagation and electric absorption (side view) (E electric field) (The value of |E|max is 1.18 × 104 V·m−1). d Magnetic field distribution during normal propagation and magnetic absorption (side view) (H magnetic field) (The value of |H|max is 3.11 × 101 A·m−1). e Contrast of motion artifact between existing modalities and the ISC scarf interface. f Contrast of noise between existing modalities and the ISC scarf interface. g Contrast of EMI between existing modalities and the ISC scarf interface (EMI electromagnetic interference) (Contains photos by Freepik via Magnific License). h Contrast of comfort between existing modalities and the ISC scarf interface. (Contains photos by studiogstock via Magnific License). The foundation of clear speech decoding lay in baseline stability and signal stability under conditions of heavy motion. Therefore, the mechanisms of generating and eliminating motion artifact were compared among the ISC scarf interface, microphones, and piezoelectric films (Fig. 1e). The former two exhibited motion artifact arising from vibrations that perturbed the microphone diaphragm capacitance and from relative sliding motion associated with adhesive skin contact (Fig. 1e-i). In contrast, the ISC scarf interface transformed conformal mechanical displacement, which originally had the same scale as human motion, into the wavelength-scale perturbation of SSP modes (Fig. 1e-ii). This scale separation dramatically suppressed motion artifact. Figure 1f compared the noise susceptibility of microphones with the ISC scarf interface. Microphones remained vulnerable to acoustic noise because diaphragm vibrations modulate the device capacitance (Fig. 1f-i). By contrast, the ISC scarf interface leveraged the linear superposition principle shared by electromagnetic and acoustic waves: while sound waves required energy transfer between media to induce mechanical motion, the ISC scarf interface relied solely on alternating electromagnetic fields, which did not undergo mechanical coupling with airborne acoustic disturbances (Fig. 1f-ii). As a result, acoustic noise did not disturb the electromagnetic sensing pathway. Another significant traditional limitation was the microphone’s low immunity to EMI, leading to the microphone’s signal distortion. The mechanisms underlying the susceptibility of microphones and the ISC scarf interface to EMI were investigated (Fig. 1g). Microphones were susceptible to external interference because incident electromagnetic fields alter the internal capacitor field distribution (Fig. 1g-i). The ISC scarf interface, however, exploited momentum-mismatch decoupling: SSP modes supported on the interface possessed substantially larger in-plane wave vectors than far-field radiation in free space, thereby violating wave-vector conservation and preventing coupling (Fig. 1g-ii). Additionally, the SSP field energy was tightly confined near the device and decays exponentially away from it, preventing spatial overlap with long-wavelength far-field interference. Wear comfort is particularly important for piezoelectric films due to their skin-adhesion sensing nature. The user comfort associated with piezoelectric films and the ISC scarf interface was compared (Fig. 1h). The former required airtight skin adhesion, which led to discomfort during prolonged wear (Fig. 1h-i), and its material choices were constrained by biocompatibility requirements. Such airtight adhesion reduced interfacial ventilation between the skin and the surrounding environment. As a result, heat dissipation and sweat evaporation were hindered, leading to local heat and moisture accumulation during extended use. In addition, continuous adhesive contact may increase skin irritation under natural body motion, further reducing wearing comfort. By contrast, the ISC scarf interface employed a textile-based substrate and a body-conformal topological embroidery structure (Fig. 1h-ii). Its breathable fibers and imperceptible sensing mechanism enabled comfortable, long-term silent speech decoding for aphasiacs (Fig. 1h-ii). The performance advantages of the system were summarized using a Sankey diagram, a table, and a multifunctional aphasiac-assisted decoding training system was subsequently developed (Supplementary Table 1 and Supplementary Figs. 2 and 3). This system integrated physiological monitoring capabilities (e.g., sleep apnea monitoring) with a decoding training assistant (e.g., silent speech recognition and robot-assistive decoding training). It operated reliably under high-noise, high-EMI, and motion-rich environments, establishing a foundation for practical deployment of ISC scarf interface-based systems. Structural design and function characteristics of the ISC scarf interface Beyond comparisons with microphones and piezoelectric films, the ISC scarf interface also demonstrated fundamental advantages in the system. To clarify the distinction between conventional systems and the ISC system, the sensing–communication separation and ISC were compared (Fig. 2a, b). In conventional systems, sensing and communication were usually implemented as physically separated modules, so the sensed signal commonly underwent an intermediate conversion and interface transfer before communication and decoding. Such structural separation increased component redundancy and made the signal more susceptible to mismatch, interference, and distortion during inter-module transduction and transmission. By contrast, the ISC presented here formed an integrated structure with coherent transduction, in which the perturbation induced by muscles was directly mapped onto the unified SSP field and transmitted through the same physical interface. This design therefore reduced the need for an additional sensing-to-communication conversion stage and suppresses interface mismatch between discrete modules. It should be noted, however, that the output signal was still acquired by an external circuit and converted from analog to digital form before backend decoding. Therefore, the advantage of the ISC architecture was not the complete removal of digitization, but the reduction of intermediate conversion steps and cumulative distortion prior to data acquisition and decoding. As a result, multi-muscle phase information was transmitted with a pristine waveform and retained for decoding. This integrated pathway enabled high fidelity while sustaining spatial selectivity, ensuring robust decoding even under motion and external disturbance. The comparison highlighted how the ISC scarf interface departs from and surpasses conventional separated implementations. To further reveal how the ISC maintains integration while balancing fidelity and dispersion-regulated selectivity, the following examined the full device design.Fig. 2Structural design and ISC characteristics of the ISC scarf interface.a Limitations of sensing–communication separation. b ISC enabled by spoof surface plasmon. c Structural components of the ISC scarf interface: (i) ISC scarf interface; (ii) Metamaterial and fiber substrate; (iii) Topological embroidery design and its anti-EMI function. d Electric field distribution of thedevice's surface mode. (S1 to S5 represent different geometric shapes.) e Normalized dispersion of different structures. (kssp absolute value of the wave vector of the SSP. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\parallel }$$\end{document}k∥ horizontal component of the wave vector; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\perp }$$\end{document}k⊥ vertical component of the wave vector; klightline absolute value of the wave vector of the light line). f Schematic diagram of actual embroidery and comparison of transmission and reflection coefficients under substrates with different permittivity. Scale bar 2 cm. g Group delay dispersion of substrates with different permittivity. h Signal transmission flow graph of the ISC scarf interface. (CNN convolutional neural network; L1, L2, and L3 represent the equivalent inductance of different device regions; Csub and Cvar represent the substrate and the varying capacitance, respectively.) i Phase response of the ISC scarf interface (Data are mean ± SD of 5% for estimated measurement uncertainty, N = 3). j Transmission mechanism of the ISC scarf interface when muscles approach or contract (t0 and t1 represent times t = 0 and t = 1, respectively) (Contains photos by Master.f.dirt via Freepik License). k Finite element simulation with and without a body (The value of |E|max is 2.83 × 103 V·m−1). l Phase change when muscles approach at different distances. Structural design and ISC characteristics of the ISC scarf interface. a Limitations of sensing–communication separation. b ISC enabled by spoof surface plasmon. c Structural components of the ISC scarf interface: (i) ISC scarf interface; (ii) Metamaterial and fiber substrate; (iii) Topological embroidery design and its anti-EMI function. d Electric field distribution of thedevice's surface mode. (S1 to S5 represent different geometric shapes.) e Normalized dispersion of different structures. (kssp absolute value of the wave vector of the SSP. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\parallel }$$\end{document}k∥ horizontal component of the wave vector; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\perp }$$\end{document}k⊥ vertical component of the wave vector; klightline absolute value of the wave vector of the light line). f Schematic diagram of actual embroidery and comparison of transmission and reflection coefficients under substrates with different permittivity. Scale bar 2 cm. g Group delay dispersion of substrates with different permittivity. h Signal transmission flow graph of the ISC scarf interface. (CNN convolutional neural network; L1, L2, and L3 represent the equivalent inductance of different device regions; Csub and Cvar represent the substrate and the varying capacitance, respectively.) i Phase response of the ISC scarf interface (Data are mean ± SD of 5% for estimated measurement uncertainty, N = 3). j Transmission mechanism of the ISC scarf interface when muscles approach or contract (t0 and t1 represent times t = 0 and t = 1, respectively) (Contains photos by Master.f.dirt via Freepik License). k Finite element simulation with and without a body (The value of |E|max is 2.83 × 103 V·m−1). l Phase change when muscles approach at different distances. As a core component of the ISC scarf interface, the metamaterial textile structure required a precise response to the subtle activities of the neck muscles. Accordingly, the interface provided SSP modes for confined field detection while maintaining excellent signal fidelity. Furthermore, the interface overcame the drawbacks associated with nonlinear phase response in dispersion environments by adopting natural and artificial polyester fibers. The structure, schematic, and geometrical design parameters of the ISC scarf interface were shown in Fig. 2c–e. The physical structure of the ISC scarf interface and its longitudinal structural decomposition, including the metamaterial layer and the substrate, were designed and demonstrated (Fig. 2c). The interface exploited the permittivities contrast between the metamaterial and the substrate to generate SSP modes at their boundary and to engineer a confined field. The relatively small but opposite numerical permittivities of the two materials, combined with the topological embroidery design, enabled artificial control over the propagation depth and modal shape of the surface wave. This confined field coordinated the propagation velocities of different frequency components, preserving the dispersion-induce confined field while supporting high-throughput, high-fidelity transmission of broadband silent speech signals. The topological embroidery design of the waveguide produced a band-flattening effect in the localized modal coupling (Fig. 2c-iii), which facilitated wavelength compression and yields spatially directional confinement of electromagnetic energy, thereby naturally enhancing immunity to external EMI. The trapezoidal and central impedance-matching sections enabled efficient wave-vector conversion between the curved plasmonic waveguide and the 50 Ω coplanar waveguide, minimizing reflections. These transition impedance-matching sections improved the transmission efficiency by more than 1.56 dB and reduced the conversion loss between the topological embroidery design and the coplanar waveguide (Supplementary Fig. 4). The SSP exhibited a wavelength compression to approximately 38.2% the free space wavelength32. The highly confined surface-wave mode and the polyester substrate provided robustness against interference sources such as large-scale human motion, acoustic noise, and electromagnetic disturbance. The SSP modes propagated in a TM polarization along the curved interface. In the x and y plane, the field energy was concentrated along the central transmission line and gradually weakens toward the curved region. In the x and z plane, the electric field extended transversely along the transmission path while decaying exponentially in the vertical direction (Fig. 2d). Here, w denoted the SSP structure width, and p denoted the SSP structure period of the structured interface. The SSP structure width w mainly determined the SSP field confinement strength and corresponding dispersion intensity, as reflected by the degree of deviation from the light line. In contrast, the SSP structure period p mainly regulated the relative linearity of the dispersion curve and therefore governed the balance between signal fidelity and dispersion. Increasing w mainly changed the confinement-related dispersion deviation, whereas varying p primarily affected the linearity of the dispersion profile while preserving the dispersion (Supplementary Fig. 5). Whether the conflict between signal fidelity and dispersion could be resolved determines whether the interface could stably transmit signals while detecting subtle local multi-muscle activities. Normalized dispersion curves were used to characterize the dispersion strength and confined field capability of the ISC scarf interface (Fig. 2e). Compared with the light line (black line), the topological metamaterial-induced dispersion exhibited a controllable deviation from the initial to the cutoff frequency, indicating that the electromagnetic energy was strongly confined at the interface. The topological design maintained a stable dispersion increase across the desired communication band of 1.5–3 GHz, demonstrating a stable mapping to the confined field region while preserving superior signal fidelity. Among the candidate geometries, S1 was selected because its dispersion curve remained relatively closer to the light line while still maintaining clear confined-field dispersion over the operating band, regulating spatial selectivity (Supplementary Note 1). This suggested a more linear phase evolution during propagation, which helped reduce waveform distortion and preserve signal fidelity, while still ensuring sufficient field confinement. In addition, the dispersion values in this frequency response range demonstrated the dispersion modulation of the ISC scarf interface. This was strong evidence that the interface has spatial selectivity. Since the distance between the light line and the interface was a dispersion value, it indictated that the SSP was generated at the surface of the ISC scarf interface. The specific wave vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\bf{k}}}}_{{SSP}}$$\end{document}kSSP satisfied the boundary conditions with its magnitude is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{SSP}}=|{{{\bf{k}}}}_{{SSP}}|$$\end{document}kSSP=∣kSSP∣, which should be equal to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{light\; line}}$$\end{document}klightline in vacuo. For the SSP to propagate smoothly and continuously on both sides of the ISC scarf interface, the wave vector component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\parallel }$$\end{document}k∥ along the interface contained in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{SSP}}$$\end{document}kSSP must be greater than or equal to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{light\; line}}$$\end{document}klightline in free space. According to the wave equation, the square of the wave vector must satisfy:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{SSP}}^{2}={k}_{\parallel }^{2}+{k}_{\perp }^{2}$$\end{document}kSSP2=k∥2+k⊥22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta \equiv {k}_{\parallel } > {k}_{{light}\,{line}}$$\end{document}β≡k∥>klightlinewhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\perp }$$\end{document}k⊥ represents the vertical component of the wave vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β represents the dispersion which equal to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}k. When the parallel wave vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\parallel }$$\end{document}k∥ was greater than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{light\; line}}$$\end{document}klightline for smooth and continuous propagation, the square of the vertical wave vector becomes negative to satisfy the wave equation, making the vertical wave vector an imaginary number:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\perp }=i\kappa$$\end{document}k⊥=iκwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\kappa$$\end{document}κ denotes a real number. Substituting this imaginary wave vector into the wave phase factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}^{i{k}_{\perp }x}$$\end{document}eik⊥x. Then the oscillation factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}^{{ikx}}$$\end{document}eikx became the exponentially attenuating factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${e}^{-\kappa x}$$\end{document}e−κx, which caused the SSP to attenuate in the vertical direction and form a confined field to resist EMI. The effects of different topological embroidery designs and single-period lengths on the normalized dispersion were evaluated (Supplementary Fig. 4). The results demonstrated that the ISC scarf interface, under varying parameters of structure, width (w = 2 mm), and period length (p = 11 mm), maintained stable dispersion across all frequencies while achieving the lowest loss relative to the light line. The degree of opposite permittivities between the substrate and the embroidery material largely determined the dispersion of the SSP. Selecting a suitable substrate and using appropriate technology to meet the basic principle of the SSP was that the permittivities of the materials on both sides of the interface were one positive and one negative. For example, the permittivity of the metal was negative at a specific frequency, and the permittivity of the adjacent dielectric was positive. Topological embroidery with different substrate permittivities was sewn, together with a comparison of their transmission and reflection coefficients, which were the parameters most directly related to SSP propagation performance (Fig. 2f). This topological embroidery structure was formed by two layers of materials with opposite dielectric constants, creating a SSP: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{{ME}}{\varepsilon }_{{SI}} < 0$$\end{document}εMEεSI<0where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{{ME}}$$\end{document}εME represents the metamaterial topological embroidery structure, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{{SI}}$$\end{document}εSI represents the polyester scarf interface substrate. For the TM surface mode, the boundary condition can gives the interface-mode dispersion as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta={k}_{{light}\,{line}}\sqrt{\frac{{\varepsilon }_{{ME}}{\varepsilon }_{{SI}}}{{\varepsilon }_{{ME}}+{\varepsilon }_{{SI}}}}$$\end{document}β=klightlineεMEεSIεME+εSI The transmission coefficient and reflection coefficient, respectively, describe how much of the surface wave propagates across the interface and how much is reflected, thereby indicating whether the wave remains effectively confined and propagation along the interface. Four substrates with progressively decreasing permittivities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{{sub}}$$\end{document}εsub were selected, and used a digital embroidery machine to fabricate the topologically woven structures, ensuring intimate contact between the metamaterial layer and each substrate (Supplementary Fig. 5). These four substrates, namely flax, cotton, non-woven fabric, and polyester, were selected because they represent common textile categories used in wearable interfaces and cover distinct fiber compositions and structural forms, including natural woven fabrics, non-woven textiles, and synthetic fabrics. This comparison, therefore, provided a representative basis for evaluating how different textile substrates affect SSP propagation. The comparison of transmission and reflection coefficients showed that the polyester substrate used in the ISC scarf interface achieved both a lower reflection coefficient and a higher transmission coefficient at the standard 2.4 GHz communication band (Fig. 2f and Supplementary Fig. 6). A higher transmission coefficient indicated that more SSP energy was maintained in forward propagation along the interface, which was beneficial for preserving the propagating signal and its phase information. In contrast, a lower reflection coefficient indicated increased interface mismatch and more backward scattering, thereby increasing reflection-induced loss and distortion. Taken together, these results indicated that polyester provided the most favorable propagation condition among the tested substrates, enabling more stable confined-field guidance with reduced scattering and reflection loss. Group delay dispersion (GDD) and group delay were key indicators of signal fidelity. The GDD and group delay in the 1.5 to 3 GHz band were evaluated for different substrates to examine how signal fidelity in the ISC scarf interface coexisted with dispersion-enabled confined field equalization (Fig. 2g and Supplementary Fig. 7). Across this band, the ISC scarf interface exhibited an overall group delay dispersion of 7 fs², indicating a highly stable dispersion profile and ultrahigh fidelity. Previously reported numerical measurements of GDD in other systems, which separated the sensing and communication, were also examined33–36. It was found that their dispersion values remained substantially higher, typically exceeding 83 to 45,000 fs², demonstrating severe distortion and an inability to achieve high fidelity and spatial selectivity. It was incompatible with the locally confined physiological communication regime explored in this work. The minimized dispersion in the design prevented severe signal distortion while still maintaining the moderate non-zero dispersion required for confined field sensing. Other substrates exhibited either large dispersion across the entire band or acceptable performance only near 2.4 GHz, but severe dispersion elsewhere. Group delay was another critical factor for high signal fidelity. ISC scarf interface demonstrated a group delay distribution centered around 5 ns, with a narrow histogram width and a variation range of less than 1 ns (Supplementary Fig. 8). To maintain high-fidelity and moderate dispersion signal performance after processing, the signal flow diagram of a 2.4 GHz carrier transmitted via universal software radio peripheral (USRP) and its processed phase were presented (Fig. 2h, i). The USRP transmitter (Tx) sent a 2.4 GHz carrier into the scarf interface, where the metamaterial and polyester substrate form a resonant LC-like circuit referenced to ground. This interface combined the functions of confined field detection of multi-muscle activities and communication between Tx and Rx, forming an ISC. Owing to its high water content and anisotropic dielectric behavior, muscle tissue acted as a lossy dielectric medium. When superficial muscles contracted or approached relative to the SSP-supported textile pathway, the confined out-of-plane electric field penetrated the tissue and induced reverse-polarization of its abundant dipolar molecules. This interaction modified the local permittivity and perturbed the spatial distribution of confined field energy, producing a measurable phase modulation characteristic of SSP–muscle coupling. In the equivalent circuit view, these permittivity perturbations manifested as the effective capacitance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{{\mathrm{var}}}$$\end{document}Cvar, which jointly determined the dispersion \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β and phase response of the interface. To quantitatively explain the spatial selectivity of the dispersion energy modulation of the ISC scarf interface, a derivation was performed from the wave vector to the dielectric constant to the effective capacitance, and then to the relationship between dispersion, phase, and spatial selectivity. (Supplementary Note 1). In short, the effective capacitance increased with the increase of the dielectric constant, which in turn increased the dispersion. Since the change in dispersion was related to the change in phase, the change in phase also increased, phase fidelity and integrity would also decrease (Supplementary Figs. 9–11). Considering the balance, this work chose a group delay dispersion value of 7 fs² as the benchmark, and based on this, discussed the spatial selectivity. This helped this work determine that the phase change should be used as a parameter for most scenarios in subsequent silent speech decoding waveforms. To facilitate understanding how dispersion-regulated spatial selectivity, parameters for spatial selectivity as a function of phase change were defined and related experiments were designed:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SS}{I}_{\phi }=\frac{\left|\Delta {\phi }_{{target}}\right|}{\left|\Delta {\phi }_{{adjacent}}\right|}$$\end{document}SSIϕ=ΔϕtargetΔϕadjacentwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\phi }_{{target}}$$\end{document}Δϕtarget and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\phi }_{{adjacent}}$$\end{document}Δϕadjacent denote the phase variations induced by the target muscle and an adjacent muscle, respectively. Selecting substrates with different permittivities further tuned this equivalent capacitance and thereby adjusted the dispersion characteristics of the SSP mode. After coupling with the muscles, the modulated carrier carried the primary multi-muscle activity information. Following down-conversion by the USRP receiver (Rx), the signal was demodulated into its original waveform plus the activity-induced component. The complex conjugate module then extracted the modulation of the muscle component within the signal’s frequency band. The complex argument (complex-to-arg) module converted this component into a phase, which was subsequently processed through frequency division and fed into the neural network module for silent speech classification. The combined use of substrate permittivity and topological embroidery enabled the SSP design to support both localized confined fields and an approximately linear phase response37,38. Within the operating band, the ISC scarf interface maintained a broadband linear phase response while achieving the moderate dispersion amplitude required at 2.4 GHz (Fig. 2i). This figure, along with Fig. 2e–g, collectively validated that the ISC scarf interface had broken the long-standing conflict between signal fidelity and dispersion. Specifically, it demonstrated a balance between ultrahigh signal fidelity represented by wideband phase integrity and moderate dispersion represented by wideband dispersion-regulated (Fig. 2c–g). By contrast, the other substrates exhibited pronounced phase truncation in parts of the spectrum and an excessively broad overall phase-response range. For example, low frequency of deviations from linearity and high frequency of jitter (Supplementary Fig. 11). After ensuring the integrity of the complete signal processing flow of the ISC scarf interface, it was necessary to investigate whether the electromagnetic changes on the surface of the ISC scarf interface during electromagnetic-muscle coupling conformed to the design (Fig. 2j, k). The electromagnetic-field transmission mechanism at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{0}$$\end{document}t0 simulated the transition of muscle tissue from a relaxed, far-field state to an approaching and contracting state (Fig. 2j). At \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{0}$$\end{document}t0, the muscle remains uncontracted, its effective thickness is unchanged, and the field interaction occurs primarily in the far-field. Under this condition, perturbation to the SSP field is minimal, producing a negligible phase change. When the muscle contracts and approaches into the near-field at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{1}$$\end{document}t1, the effective penetration depth of the electromagnetic field decreases, while the tissue area engaged by the confined field increases. A larger fraction of the field that would otherwise propagate unimpeded is absorbed by the muscle, attenuating the confined electromagnetic energy. This attenuation manifests as a phase change in the SSP, which was subsequently converted into usable information through the signal processing. The electric field energy distribution at a specific phase and its variation as the muscle approached the interface were demonstrated (Fig. 2k). The field distribution was consistent with the proposed mechanism: energy gradually increases at the impedance-matching nexus, and the topological embroidery design directed the SSP energy uniformly along the central transmission line. Electromagnetic absorption simulations of biological tissue further confirmed that muscle absorption maps directly to phase modulation of the confined SSP field. Supplementary Video 1 provided the complete temporal evolution of the field distribution during propagation. To visualize the change in electromagnetic energy, the original signal distributions at distances of 6, 3, and 1 cm between muscles and the interface were shown (Fig. 2l). As the muscle approached the interface, the peak amplitudes of the two waveforms decreased markedly, consisting with the mechanisms of the ISC scarf interface and with finite element simulations showing progressive absorption of electromagnetic energy by the muscle. Robust body sensing with the ISC scarf interface The most pressing need for aphasiacs is reliable silent speech decoding in the real-world as well as features such as portability and flexibility, like electronic skin and wearable devices39–41. Therefore, the ISC scarf interface required the successful establishment of a complete muscle-to-electromagnetic-to-phase-to-speech mapping chain. To create a direct correspondence between articulatory muscle and silent speech, side-view and front-view postures of representative international phonetic alphabet (/ɑː/, /ɔː/, /ə/, /iː/) during silent speech were captured (Fig. 3a and Supplementary Fig. 12). Figure 3b showed the specific absorption rates (SAR) of the laryngeal and mandibular regions. The silent speech region, primarily composed of bilateral laryngeal muscles, exhibited SAR values of 0.06–0.184 W kg−1, whereas the mandibular region presented 0–0.06 W kg−1. This difference indicated that the propagating SSP mode coupled more strongly to the laryngeal region than to the mandibular region under the same excitation condition. Notably, the peak SAR in the laryngeal region (0.184 W kg−1) was about 306% higher than the upper bound of the mandibular region (0.06 W kg−1), indicating a substantially stronger local electromagnetic interaction. In this context, the SAR reflected the spatial strength of SSP energy absorption and therefore the local sensitivity of the interface to tissue perturbation, rather than the muscle activity intensity itself. As a result, muscle contractions occurring in the higher-SAR laryngeal region were expected to induce larger perturbations to the SSP propagation state, leading to more evident phase modulation than those arising from the lower-SAR mandibular region. The SAR distribution, therefore, provided a quantitative physical explanation for the spatial selectivity of the ISC scarf interface and its ability to distinguish laryngeal and mandibular configurations even in the absence of audible speech. The SAR distribution also confirmed that the ISC scarf interface met the IEEE C95.1 safety standard for both aphasiacs and the general population. To illustrate the mechanism underlying information integration, the complete activities-induced phase modulation process was mapped as the SSP propagated along the neck regions (Fig. 3c). By mapping phase changes to the distribution changes of power absorption, and then inferring the activities patterns of different muscles, this further verified that the ISC scarf interface resolved the intrinsic conflict between dispersion and fidelity, and achieved spatial selectivity.Fig. 3Robust performance and mechanisms of body sensing on silent speech.a Physical display of different international phonetic alphabets after wearing the ISC scarf interface (Participants; sex male; number 1; age 27). b Specific absorption rates of the laryngeal and mandibular muscles (SAR specific absorption rates). c Establishment of direct mapping between muscles, fields, and phases. d Electromyography reference for phonetic alphabets. e Original waveform of phonetic alphabets. f LDA analysis between electromyography sensor and the ISC scarf interface (LDA linear discriminant analysis). g Comparison of physiological assessment curves for the ISC scarf interface and existing modalities (Contains photos by ssstocker via Freepik License). From left to right, these represent daily physiological activities: swallowing, yawning, shaking, talking, and making phone calls. Curve positions: microphone (top), piezoelectric film (center), and wearable platform (bottom). h Bee swarm plot of different international phonetic alphabets. i Comparison of curve waveforms of phonetically-alike terms. Robust performance and mechanisms of body sensing on silent speech. a Physical display of different international phonetic alphabets after wearing the ISC scarf interface (Participants; sex male; number 1; age 27). b Specific absorption rates of the laryngeal and mandibular muscles (SAR specific absorption rates). c Establishment of direct mapping between muscles, fields, and phases. d Electromyography reference for phonetic alphabets. e Original waveform of phonetic alphabets. f LDA analysis between electromyography sensor and the ISC scarf interface (LDA linear discriminant analysis). g Comparison of physiological assessment curves for the ISC scarf interface and existing modalities (Contains photos by ssstocker via Freepik License). From left to right, these represent daily physiological activities: swallowing, yawning, shaking, talking, and making phone calls. Curve positions: microphone (top), piezoelectric film (center), and wearable platform (bottom). h Bee swarm plot of different international phonetic alphabets. i Comparison of curve waveforms of phonetically-alike terms. The silent speech was next validated against simultaneously acquired surface electromyography (EMG) signals and the phase of the ISC scarf interface (Fig. 3d, e). The two-channel EMG traces confirmed that each phonetic alphabet evoked discriminable activation profiles, with temporally localized bursts corresponding to the onset of specific silent speech gestures. The waveform diagram of the phonetic alphabet in EMG provided a reference benchmark for subsequent ISC scarf interface measurements. Building upon this physiological grounding, the wearable scarf interface transduced muscle perturbations of the near-field electromagnetic environments into measurable phase waveforms (Fig. 3e). Four representative phonetic alphabets yielded distinct phase trajectories, each exhibiting characteristic inflection segmented that served as robust identifiers for phonetic alphabet classification. The differences between these robust identifiers stemmed from the different muscle dominance of the neck regions in silent speech. At this point, the spatial selectivity of the ISC scarf interface captured the dominant muscle activity and precisely converts it into phase information. Consequently, this muscle, field, and phase mapping substantiated another core of the ISC scarf interface that silent speech in aphasiacs could be captured and discriminated without acoustic or piezoelectric, thus enabling a silent speech interface that bypasses the vocalization entirely. The LDA clustering analysis was used to illustrate the correlation between the reference EMG signals and the ISC scarf interface signals, as well as the clear separation between different alphabets (Fig. 3f). The four alphabets showed obvious clustering within their respective clusters, indicating that the ISC scarf interface corresponds well to the EMG signals for each of the four alphabets. Furthermore, the cluster separation demonstrated excellent separation between different alphabets without affecting the waveform recognition of other alphabets. Furthermore, the phase data of the interface matched the benchmark, enabling recognition. Due to the need for physiological movement, a complete verification of the bending and twisting performance of the ISC scarf interface was essential. Its bend and twist transmission coefficients were illustrated (Supplementary Fig. 13). When the angle was in the 0°–90° range, which conformed to the human neck, the transmission coefficient remained high, at −3 to −4 dB (Supplementary Fig. 13). The bending angle of a scarf when worn fell within this range. When the ISC scarf interface was tested for electromagnetic wave absorption by the human body within this frequency and bending angle range, it was found that the transmission coefficient was not affected by the proximity of the human body at the desired communication frequency. In real life, scarf interfaces should not only conform to the human body but also account for the random twists caused by human movement. The transmission coefficient of the scarf interface, from initial flatness to extreme twist, ranges from −2 to −4 dB. In addition, cyclic bending tests further confirmed the performance stability of the ISC scarf interface, as the normalized phase response remained nearly unchanged after 10,000 bending cycles, with no evident drift or waveform degradation (Supplementary Fig. 14). This meant that its transmission performance remained stable even in these multiple bent and twisted states, adapting to some unpredictable human behaviors. To eliminate interference from other physiological activities in silent speech and compare, the ISC scarf interface continuously monitored the same physiological assessment from the microphone to the piezoelectric film to the ISC scarf interface: swallowing, yawning, random body movement, silent speech, and phone calls (Fig. 3g). While the microphone was clearly insensitive to swallowing and yawning, it inevitably exhibited significant, high noise during random body movement and phone calls, along with low repeatability of the speech waveform. Due to the high sensitivity of the electromechanical mechanism to motion, the piezoelectric film maintained high baseline jitter throughout the entire process. The shortcomings of traditional rigid substrates, microphones, and piezoelectric films for detecting physiological activities in the human body were also analyzed (Supplementary Fig. 15). Various motion artifact caused jitter on the piezoelectric film, even up to 1237% of the initial value. Compared to the ISC scarf interface, the piezoelectric film’s motion artifact significantly interfered with signal stability (Supplementary Fig. 16). While the repeatability of silent speech recognition was high, significant interference signals were also observed when subjected to EMI. The ISC scarf interface demonstrated clear amplitude in common physiological activities, such as swallowing and yawning, indicating that even subtle muscle contractions can trigger device feedback. However, when it came to motion artifact and EMI, which could be detrimental to the signal, the scarf interface smoothly transitioned through motion artifact and more intense movements, with an average jitter of 5% relative to the initial signal (Supplementary Fig. 15). Importantly, the ISC scarf interface also demonstrated significant discrimination when used to recognize simple words. Research on motion artifact reduction was also investigated42–45. The survey results showed that most wearable devices reduced motion artifact by only 200 to 500%, far less than the 24,300% reduction achieved by the ISC scarf interface, demonstrating the significant performance of the ISC scarf interface in reducing motion artifact. Moreover, anti-EMI research was investigated46–49. Even the best-performing anti-EMI was only around 38 dB, which is 591.8% worse than the anti-EMI of the ISC scarf interface. The statistical metrics for the robustness of these two important ISC scarf interface environments and humans were also validated (Supplementary Note 3, Supplementary Table 2 and Supplementary Fig. 17). To illustrate the high robustness and signal similarity of the ISC scarf interface, a normalized cross-correlation test and short-time Fourier transform curve were performed on the physiological assessment (Supplementary Fig. 18). Even when the microphone exhibited poor signal robustness, such as in Movement 1 and Movement 6, the ISC scarf interface maintained high similarity in physiological signal monitoring. However, the piezoelectric film exhibited significantly higher performance in certain movements, such as movements 4 and 5. This did not change its unsuitability for achieving stable similarity in most physiological signals. Compared to the instability of the piezoelectric film and the low robustness of the microphone, the ISC scarf interface maintained a consistently high level of normalized cross-correlation overall. It also maintained similarity and repeatability in certain situations where both the microphone and the piezoelectric film exhibited low robustness, such as movements 1, 2, 3, and 6. Full verification of the ISC scarf interface’s speech decoding capabilities required simultaneous recognition of vowels and consonants, as well as some phonetically-alike words that were hard to recognize (Fig. 3h, i). Bee swarm plots of the phase distribution of vowels and consonants were used to verify the ISC scarf interface’s ability to discriminate finely across most phrases (Fig. 3h). There was a strong correlation between the mouth opening amplitude and the muscle contraction amplitude of vowels. The former two vowels /ə/ and /ɑ:/ had significant phase changes. Although they used similar muscles during articulation, with the masseter being dominant (the jaw pushes forward), the swarm phase distribution diagrams of these two vowels over multiple movements reflected their recognition differences, and there was essentially no problem of confusion. The latter two consonants, compared to the former two vowels, involved the amplitude of the thyrohyoid muscle (pulling the laryngeal downward). The amplitude of the thyrohyoid muscle pulled on the latter two consonants (/s/ and /i:/) was significantly less than that on vowels, but they can still be clearly distinguished by the ISC scarf interface. Furthermore, Supplementary Fig. 19 illustrated the differences between some vowels and consonants when the muscles in the laryngeal and mandibular regions were dominant during silent speech. To evaluate the practical significance of multi-phrase recognition, the phrases with similar pronunciations using the ISC scarf interface was further examined (Fig. 3i). These phrases were intentionally selected as a representation of easily confused words, not only because they were close in muscle activity, but more importantly because they probed whether the ISC scarf interface could resolve subtle yet functionally meaningful differences in silent speech. First, similar phrases often recruited partially overlapping muscle groups while differing in the contraction of the laryngeal and mandibular muscles. Therefore, successful discrimination of such phrases indicated that the interface was sensitive to physiologically distinct motor patterns rather than merely capturing coarse speech categories. Second, in practical silent-speech communication, confusion most often arose between words or phrases that were articulatorily similar rather than obviously different. Thus, distinguishing these near-confusable examples was a more relevant benchmark for real-world usability than distinguishing phrases with large articulation differences. For example, phrases such as Chong Qing and Mao Bing, although sharing similar rhyming characteristics, involved different throat–jaw coordination patterns. The corresponding waveform differences suggested that the interface could capture distinct combinations of laryngeal and mandibular activity. More notably, for pairs such as Shi Shi and Si Shi, whose pronunciations and overall contraction patterns were highly alike, the main difference was concentrated in the initial consonant articulation (Supplementary Fig. 20). The ability to resolve such minimal articulatory distinctions demonstrated that the ISC scarf interface could detect fine variations in muscle activity. This capability was important not only for recognizing phonetically similar words, but also for supporting more reliable silent-speech decoding in realistic usage scenarios where many candidate words shared closely related articulation patterns. Robust operation in extreme environments and physiological monitoring After verifying the silent speech sensing of the ISC scarf interface and its advantages in diverse physiological activities, its performance in real-world applications needed to be explored just like the environmental adaptability verified by some non-contact wearable fabric devices50. Figure 4 showed the robustness and advantages of the scarf interface in multiple applications, demonstrating the device’s potential for translation from physical mechanisms to practical applications. Thus, the effect of an aphasiac wearing the ISC scarf interface communicating versus a bystander in a noisy and highly EMI environments were shown (Fig. 4a). Communicating in such a noisy environment was extremely difficult for normal people. While aphasiac wearing the ISC scarf interface could produce clear speech simply by silently contracting neck muscles and was unaffected by external disturbances. Moreover, the ISC scarf interface allowed the aphasiacs to communicate and make requests awake, while also providing some monitoring and support for their health even when being unconscious (Fig. 4b). In sleep apnea, the unique phase waveforms were allowed to verify different types of sleep apnea to prevent potential health risks and verify the muscle tone of silent speech after. Neck-related breathing muscles exhibited characteristic tension patterns during sleep and were used as biomarkers in distinguishing obstructive sleep apnea (OSA) from central sleep apnea (CSA). In the CSA, the absence of breathing drive produced a flat signal, whereas the OSA generated small fluctuations due to residual but ineffective inspiratory efforts, and normal breathing yields large periodic modulations. These well-established muscle-tension signals demonstrated that subtle muscles can be reliably captured even under low-amplitude and complex environments. Because the articulatory muscles used for attempted vocalization in aphasiacs largely overlapped anatomically with these neck groups when breathing, the similar activity patterns ensured that the subtle multi-muscle activities underlying silent speech can be used to distinguish sleep apnea. For example, the CSA indicated the potential risk of heart disease, while the OSA indicated the risk of sleep-deprived driving fatigue, stroke, and other consequences. Therefore, specific performance demonstrations were needed to verify performance in extreme environments and real-world applications. The performance advantages of the ISC scarf interface under extreme environments were demonstrated using the waveforms of eight phrases detected under extreme disturbance (Fig. 4c). From top to bottom, the six curves were for the piezoelectric film, microphone, and the ISC scarf interface. The piezoelectric film, under the EMI of 233 V/m and the noise interference of 85 dB, no longer displayed a clear phrase waveform. The microphone also exhibited some background noise due to noise interference. This contrasted with the clear waveforms from the ISC scarf interface. When aphasiacs talked silently, the waveforms of eight common physiological needs and requested for help were captured and identified with remarkable clarity and recognition (Supplementary Video 2 and Supplementary Fig. 21). The power spectrum and SNR of the SSP near the ISC scarf interface main frequency were used to illustrate its recognition performance in extreme environments (Fig. 4d). The signal bandwidth was strictly set to 0.05 kHz around the main frequency of 0.025 kHz, achieving an SNR of 33.44 dB. This signal range precisely encompassed the muscle contraction frequency under silent speech, and the SNR of this frequency band strongly demonstrated the ISC scarf interface’s ability to capture muscle contractions. A series of studies on the SNR of wearable biosensing were also investigated to demonstrate the superiority of the ISC scarf interface51,52.Fig. 4Multiple applications and physiological monitoring.a Concept of silent speech in dynamic interference environments. (Contains photos by jevaistuvais and yummybuum via Freepik and Magnific License). b Concept of sleep apnea monitor with the ISC scarf interface. (CSA central sleep apnea, OSA obstructive sleep apnea) (Contains photos by brgfx, Freepik, and witsanu_sinlapachai via Magnific and Freepik License). c Robustness comparison of different phrases and devices under high noise and EMI. d Power spectrum and SNR of the ISC scarf interface at the dominant frequency (SNR signal-to-noise ratio). e Real display of complex environments phrase detection using the ISC scarf interface (Participants; sex male; number 1; age 27). f Apnea detection under different sleeping postures (Contains photos by Freepik via Magnific License) g Differentiation curves between CSA, OSA, and MSA (MSA mixed sleep apnea). h Practical demonstration of the distinction between CSA and OSA. i Prolonged wearing comfort and no allergy (Participants; sex male; number 1; age 27). j High breathability of ISC scarf interface. k Steady temperature after 24 h continuous use. Multiple applications and physiological monitoring. a Concept of silent speech in dynamic interference environments. (Contains photos by jevaistuvais and yummybuum via Freepik and Magnific License). b Concept of sleep apnea monitor with the ISC scarf interface. (CSA central sleep apnea, OSA obstructive sleep apnea) (Contains photos by brgfx, Freepik, and witsanu_sinlapachai via Magnific and Freepik License). c Robustness comparison of different phrases and devices under high noise and EMI. d Power spectrum and SNR of the ISC scarf interface at the dominant frequency (SNR signal-to-noise ratio). e Real display of complex environments phrase detection using the ISC scarf interface (Participants; sex male; number 1; age 27). f Apnea detection under different sleeping postures (Contains photos by Freepik via Magnific License) g Differentiation curves between CSA, OSA, and MSA (MSA mixed sleep apnea). h Practical demonstration of the distinction between CSA and OSA. i Prolonged wearing comfort and no allergy (Participants; sex male; number 1; age 27). j High breathability of ISC scarf interface. k Steady temperature after 24 h continuous use. To eliminate movement interference during sleep apnea, normalized curves and images of sleep apnea monitoring in different sleeping positions were demonstrated (Fig. 4f and Supplementary Fig. 22). Lying on the left or right side compressed the ribs, limiting breathing space in the ribs. The amplitude of the phase curve during sleep was consistently smaller than that of lying on the back. During this process, curves for different breathing rates and for significant recovery after apnea were also monitored. This also provided various auxiliary curves for identifying sleep apnea and other risks in aphasiacs, such as whether they experienced shortness of breath or hyperpnea. Physiological health monitoring of aphasiacs was implemented using the ISC scarf interface by classifying sleep apnea curves (Fig. 4g). Both the CSA and OSA experienced apneas, but the mechanisms of these apneas differed. In the CSA, the breathing drive was absent or abnormal, preventing the brain from effectively transmitting breathing commands. In the OSA, breathing tract obstruction persisted, but breathing drive and effort were present. Consequently, during the OSA, the laryngeal and abdomen exert a struggle to apnea, similar to the struggle of suffocation. Therefore, compared to the complete absence of a resistance curve in the CSA, the OSA exhibited a resistance curve that is lower than the breathing amplitude during normal breathing (Fig. 4h and Supplementary Video 3). The video demonstrated switching between normal breathing, CSA, normal breathing, OSA, and normal breathing, showcasing the ISC scarf interface’s application in a versatile physiological indicator monitoring system (Supplementary Fig. 23). This reflected its potential for real-world applications, from extreme environments with complex interference to subtle sleep apnea monitoring. Because the ISC scarf interface needed to be worn for extended periods of time in the aphasiac’s neck area, long-term skin compatibility, breathability, and constant temperature (Fig. 4i–k) are important. The ISC scarf interface did not cause any allergic reactions in aphasiac after prolonged contact with skin in a dynamic environment (Fig. 4i). Furthermore, water vapor transmission was used to test the breathability of the ISC scarf interface (Fig. 4j). From the beginning of boiling to 1 h after boiling, it was clearly that the area through which water vapor permeates was roughly the same size as the beaker opening and covered most of the interface. Turning to the constant temperature metric (Fig. 4k), a heat map analysis of the area after 24 h of signal transmission was performed, revealing that the temperature variation was approximately negligible. Similar additions were made regarding long-term breathability and temperature stability (Supplementary Fig. 24). All three of these characteristics met the human comfort requirements of a steady temperature, breathability, and comfort to the ISC scarf interface. Interactive feedback and assistant during silent speech decoding Accurate language recognition and feedback for aphasiacs was not only a key reflection of the multifaceted care provided to aphasiacs, from physiological needs to health monitoring. It was also a key foundation for their healthy integration into the social environments and the reduction of psychological illnesses caused by social barriers. In the process of social interaction, accurate interaction meant accurate interpretation and timely analysis of silent speech intent through machines learning6. Human-machine interaction-based decoding training assistant was used to illustrate how the ISC scarf interface processed and implemented feedback and assistive interaction to promote the decoding training (Fig. 5a and Supplementary Fig. 25). Based on the waveform diversity brought by multi-muscle activities to the SSP, a high sampling rate of 65 kHz was designed to prevent the omission of details. However, it was very difficult to use such high sampling rate data for processing, training, and display. This was also too high a load for a portable and lightweight device such as the ISC scarf interface. Therefore, some signal processing techniques were employed in the signal processor, such as down sampling and Fourier transform, to down sample and separate the desired muscle contraction frequencies. These different muscle contraction frequencies corresponded to the spatial selectivity resulting from dispersion-regulated, which aligned with the principle of dispersion, namely frequency selectivity. Multi-muscle activities at these frequencies essentially covered the activity frequencies of the neck region. Subsequently, the activities of these muscles were mapped to different features. These features were then fed into the convolutional neural network, which performs a 3.7 s model updating to adapt to the user’s multi-muscle activities. Through extensive evaluation and comparison with state-of-the-art silent speech systems, the optimal neural network selection for this environment was identified (Supplementary Fig. 26). The updated model reliably decoded silent speech and successfully recognized the intended command “Water” with high recognition. Through this recognition, the system generated feedback to guide the interactive decoding training assistant. Such positive feedback reinforced correct articulatory attempts and facilitated the gradual re-engagement of confidence for the aphasiac. Over repeated sessions, this closed-loop sensing–decoding–feedback process promoted muscular coordination in the neck region, thereby accelerating accurate decoding training assistant.Fig. 5Assistive interaction and training during silent speech decoding.a Framework of decoding training when interacting with machine (Contains photos by brgfx, Freepik, and cartoonsolutions via Magnific and Freepik License). b T-distributed stochastic neighbor embedding (t-SNE) two-dimensional visualization of five types of phrases (BB bye bye, GS grab something, SH shake hand, HD help dress, PS pass something, t-SNE t-Distributed Stochastic Neighbor Embedding). c Confusion matrix of five phrases recognition. d Training and validation performance of phrases recognition. e Concept of decoding training for aphasiac (Contains photos by brgfx, ilonitta, jehsomwang, and cartoonsolutions via Magnific and Freepik License). f Ablation experiment performance under different configurations. g Physical image of human-machine interaction feedback for aphasiac. h Principal component analysis curves and recognition feedback for different phrases (Participants; sex male; number 1; age 27). Assistive interaction and training during silent speech decoding. a Framework of decoding training when interacting with machine (Contains photos by brgfx, Freepik, and cartoonsolutions via Magnific and Freepik License). b T-distributed stochastic neighbor embedding (t-SNE) two-dimensional visualization of five types of phrases (BB bye bye, GS grab something, SH shake hand, HD help dress, PS pass something, t-SNE t-Distributed Stochastic Neighbor Embedding). c Confusion matrix of five phrases recognition. d Training and validation performance of phrases recognition. e Concept of decoding training for aphasiac (Contains photos by brgfx, ilonitta, jehsomwang, and cartoonsolutions via Magnific and Freepik License). f Ablation experiment performance under different configurations. g Physical image of human-machine interaction feedback for aphasiac. h Principal component analysis curves and recognition feedback for different phrases (Participants; sex male; number 1; age 27). To demonstrate the clear separation of the five phrases in terms of silent speech, a two-dimensional t-SNE plot of the five phrases was shown (Fig. 5b). Within the limited x and y axis ranges, each of the five phrases had a defined dominant distribution without confusion. The confusion matrix also illustrated the recognition performance of the five phrases under this framework, with recognition generally ranging from 96 to 100% (Fig. 5c). This had a certain degree of error, but overall, it accurately recognized many phrases. The training accuracy and training loss function represented the performance of the recognition model (Fig. 5d). To demonstrate the scalable learning capabilities of the ISC scarf interface on silent speech words and sentences, leave-one-subject-out experiments were conducted on 10 subjects and added 4000 additional samples to illustrate the interface’s recognition performance and generalization across different subjects and multiple sentence types (Supplementary Table 3 and Supplementary Fig. 27). Here, cross-entropy loss with label smoothing was adopted, and several commonly used alternative loss functions were further compared (Supplementary Fig. 28). Before epoch 50, training on all data may reach a local optimum, but after a few epochs, this optimum was broken. It was not until 75 epochs that the model’s accuracy and loss stabilized. The accuracy of the interface reached an extremely high level of 98.7%. It also achieved a rapid training time of 0.02 s per epoch, which was extremely significant for initial stabilization and improving accuracy in machine learning: it could achieve high-accuracy recognition for a model in just over 3 s. Due to the diversity of phrases required for silent speech recognition for aphasiacs and the need for multilingual adaptation, it was crucial to add required training data in real-time, achieve short training times, and improve model accuracy. The interface also conceived a human-machine interactive decoding training assistant system for aphasiacs (Fig. 5e). A medical professional or device presented a vocabulary to guide training, which the aphasiac then received through the ISC scarf interface. When the correct phrase was spoken, the robotic arm provided correct feedback and satisfied the aphasiac’s needs. This provided significant physiological satisfaction and support to the aphasiac, encouraging them to continuously practice correct phrase pronunciation while preventing them from losing confidence due to traditional, inefficient training feedback. This configuration achieved 98.7% accuracy and a training time of 3.7 s, and was used as the baseline. Notably, although reducing the model depth to two layers resulted in a clear drop in accuracy (Fig. 5f), this should not be interpreted as evidence that lightweight models performed poorly in silent speech decoding. Instead, it indicated that an excessively shallow architecture was insufficient for this task (Supplementary Note 2). Ablation experiments were used to verify the effects of different hyperparameters on the silent speech decoding system and to find a suitable silent speech decoding model configuration (Fig. 5f). Batch size referred to the number of samples involved in the forward and backpropagation computations. While a larger batch could improve the stability of gradient updates, it also increased the latency and resource demand of model updating for silent speech decoding, which was unfavorable for rapid subject-specific adaptation. In this configuration, with a batch size of 16, the accuracy gap with 64 did not improve or decrease significantly, but training time increased by nearly three times, which contradicted the goal of low-latency decoding. The number of convolutional layers represented the depth of the network and the number of convolutional operations. In this training, 2 layers showed a significant performance drop compared with 3 or 4 layers, indicating insufficient capacity to capture subtle and hierarchical muscle-activation patterns related to articulation, whereas 4 layers may introduce redundancy, longer training, and a higher risk of overfitting to subject-specific variations in the patient dataset. But it could also be temporarily invoked when an increase in recognition accuracy was needed. Similarly, the number of features, representing the number of channels in each convolutional layer, reflected the model capacity for learning discriminative decoding representations from weak silent-speech signals. Doubling the number of features significantly increased the training time, while halving the number of features made it difficult to preserve sufficient articulatory and linguistic information for accurate decoding. Weight decay applied an L2 penalty to weights and was also used to prevent overfitting. High and low weight penalties had little impact on model fit, demonstrating a high degree of fit for the dataset. This ablation experiment demonstrated that although a substantial increase in model complexity could hinder efficient training, moderate reductions in complexity and regularization did not significantly impair silent-speech decoding performance while reducing training time. This was consistent with the need for lightweight models and rapid subject-specific calibration in adaptive silent speech decoding. Figure 5h showed a physical illustration of the human-machine interaction for aphasiacs’ silent speech. When the aphasiac issued the command SH (Shake hand), the ISC scarf interface provided feedback on the correctness of the silent speech and directed the robotic arm to assist the aphasiac. The excellent feedback and interaction with the robotic arm provided significant training feedback for the aphasiac and visualized the training process. The principal component analysis of the feedback from different word phrase recognition results were used to demonstrate our aphasiac training feedback (Fig. 5i). The five phrases were SH (shake hand), GS (grab something), PS (pass something), HD (help dress), and BB (bye bye). These phrases constituted some of the physiological and communication needs of aphasiacs in their daily lives. The curves clearly reflected the distinct characteristics between different word phrases and the accurate operation results of the robotic arm. Supplementary Video 4 and Supplementary Figs. 29 and 30 illustrate this process, as well as the frequency division curves of muscle contraction frequency detected by the ISC scarf interface when saying the five different phrases. This well reflected the success of the training feedback, and that during silent speech, the region of the neck muscles only contracted at low frequencies, and was largely unaffected by high frequencies. The study demonstrated the robust silent speech decoding across held-out subjects and progressively more complex command settings, scaling the model to substantially multilingual scenarios remain challenging. The main difficulties were expected to arise from vocabulary expansion, greater articulatory and temporal variability, and cross-language distribution shifts in silent articulation patterns. Future extensions may be supported by transfer learning from language-specific pretrained models, hierarchical coarse-to-fine decoding, and incremental adaptation strategies that enable efficient expansion to new vocabularies and additional languages without retraining the full model from scratch. Discussion In this work, we presented a transformative breakthrough in electromagnetic physiological communication through an ISC scarf interface that imperceptibly captures subtle multi-muscle activities to decode silent speech ultrarobustly. Grounded in the physical transduction whereby multi-muscle activities perturb SSP fields, the system demonstrated that muscle–phase characterization can serve as a stable, high-information carrier for physiological communication. By integrating SSP-driven electromagnetic physiological communication with metamaterial topological embroidery, the textile resolved the intrinsic dispersion–fidelity conflict and maintained ultrahigh fidelity (7 fs2) under multi-muscle activities. This combination enabled local muscular dynamics, otherwise inaccessible to microphones or piezoelectric films, to be precisely encoded into the SSP phase, thereby establishing a direct route from dispersive surface-wave physics to high-fidelity physiological decoding. Moreover, the ISC scarf interface exhibited strong robustness against motion artifacts, long-duration usage, and external disturbances (24,300%; 24 h; 33.44 dB), while preserving reliable physiological information during extended wear and supporting broader assistive and health monitoring functions. The interface also enabled sleep apnea monitoring and supported assistive interaction for decoding training in aphasiacs. The system thereby extended the utility of electromagnetic physiological communication from silent speech decoding to broader myogenic interfacing and continuous health assessment. By translating multi-muscle activities into digital commands, the system provided an intuitive platform for daily assistance, with further potential for early dysphagia detection, laryngeal hypomobility assessment, and neuromuscular monitoring. Looking forward, this work points toward textile-based interfaces that integrate metamaterial electromagnetics and human electrophysiology to create highly informative, near-body sensing systems. Such electromagnetic physiological communication frameworks may support future applications in rehabilitation, continuous health assessment, and naturalistic human-machine interaction. Methods Fabrication of ISC scarf interface Using 280 D silver sewing thread (Dongguan Shengxin Special Rope Co., Ltd., 2.4–3.0 Ω·cm−1, silver-plated finished product 310–250 D, elongation 10%) as the top thread, the sewing was performed on a Wilcom-designed pattern using a fully automatic digital embroidery and sewing machine NV2700 (Ningbo Youkemei E-commerce Co., Ltd.). The base material was polyester fiber. Ordinary household sewing thread was used as the bottom thread, resulting in a ISC scarf interface. Transmission process of ISC scarf interface The ISC scarf interface uses an USRP B210 RF development board (Guangzhou Yuanxing Technology Co., Ltd.). Hardware-wise, the Tx and Rx ports are connected to the two ends of the ISC scarf interface device using SMA coaxial cables (Beiji Electronics Co., Ltd.). Using the GNU Radio Companion software 3 on an Ubuntu 18.04 system, a 2.4 GHz carrier wave is transmitted to the RF development board through the ISC scarf interface. The signal, after down-conversion by the carrier and the USRP source, undergoes complex conjugate multiplication, then passes through a Complex to Arg and bandpass filter to obtain the original phase signal with directly observable phase changes. Characterization and measurement Finite element simulations and SAR experiments of electric field absorption and ISC scarf interface transmission were conducted using CST Studio Suite software 2024 (Dassault Systems). Communication and electromagnetic wave performance experiments were performed using a Keysight E5071C 8.5 GHz four-port (1. Maximum frequency: 8.5 GHz 2. Harmonic: −25 dBc 3. Number of built-in ports: 4 ports 4. Maximum output power: 10 dBm 5. Maximum input power: 10 dBm) advanced vector network analyzer. The E5071C was used to detect dispersion relations of different structures, GDD relations of different substrates, linear phase response (Error bars of 5% for estimated measurement uncertainty), and Smith charts for impedance-matching, as well as SNR measurements. An electromagnetic radiation detector (Delixi Electronics Co., Ltd.) was used to quantify radiation values in extreme environments. Electromyography (EMG) signals from the larynx and mandible were measured using a Cheez sEMG wet electrode sensor (Qisheng Electronics Co., Ltd.) and an Arduino UNO R3 as a correlation reference for the ISC scarf interface. Signal processing and neural network usage The ISC scarf interface enables real-time signal push across software interfaces via the ZMQ PUB sink module in GNU Radio Companion 3. It utilizes Python’s deque (Python 3.11) sequence to provide parallel storage and display caching for real-time signal processing and display of large datasets. During processing of Matlab R2022a, it performs gradient down-sampling filtering and envelope calculation on the signal, reducing the sampling rate to 200 Hz. We use a three-layer convolutional neural network model and add operations such as differencing, moving average features, Gaussian noise, and random masking to enhance the model’s robustness and prevent overfitting. We truncate the training in advance during fitting by using log files and model judgment, and save the results. Application interface for aphasiac The interface for the aphasia application was edited using the QtWidgets module in PyQt5, clearly displaying the waveform on the left and the waveform recognition results on the right. The word recognition results are also displayed in the middle right. Additionally, a reset button is included in the lower right corner for easy recording, and a zoom function is enabled in the middle of the waveform, allowing medical staff or aphasia users to reset records and zoom in/out on the waveform. Statistical analysis Error bar analysis in this work represents mean ± standard deviation, showing central tendency of the data. Box plots analysis in this work were used to represent the distribution of the data. The central line in each box represented the median, the lower and upper edges of the box indicated the 25th and 75th percentiles, respectively (interquartile range, IQR), and the whiskers extended to the minimum and maximum values within 1.5 × IQR from the quartiles. Unless otherwise specified, all statistical samples in this experiment were derived from independent material samples or tests. Reporting on sex and gender The muscle groups in the neck involved in silent speech or internal silent speech are the same in both sexes. Anatomically and functionally, the core muscular systems controlling silent speech are homologous and similar, not separate mechanisms for both sexes. The only difference is the larynx and the vocal cords, but this applies to spoken speech scenarios and is not applicable to the silent speech scenarios in this research. A total of five male subjects participated in the experiments for anti-motion artifact and anti-EMI. The neural network training experiment involved a total of 10 male subjects. Participants’ gender was based on self-report. All participants were informed with written consent. The authors affirm that human research participants provided informed consent for publication of the images in Figs. 3a, 4e, h, i, and 5g, h and Supplementary Figs. 3, 12, 21, 22, 23, and 29. The person photographed in Supplementary Figs. 21, 22, and 29 is one of the authors. Blinding Blinding was not considered relevant to this study because the experiments did not involve treatment or control group allocation. All participants completed predefined silent-speech tasks and motion or EMI conditions, and the primary outcomes were objective, algorithm-derived metrics, including recognition accuracy and statistical measures of motion-artifact robustness and EMI robustness. Recognition accuracy was calculated by comparing model outputs with predefined ground-truth labels, and motion-artifact-related and EMI-related analyses were performed using signal-processing and statistical procedures. These results are based on the sensor’s output to the human body and are not related to the researchers’ subjective judgment. Therefore, the investigator’s awareness of the participant’s identity or experimental condition was unlikely to introduce subjective assessment bias. Sample size rationale No formal statistical sample size calculation was performed because this study was designed as an exploratory engineering validation of a non-contact scarf-based silent-speech sensing system rather than a hypothesis-driven clinical or treatment-control study. The sample size of 10 participants was chosen pragmatically to provide preliminary cross-participant validation while maintaining a feasible repeated-measures data-collection protocol. Each participant completed the same predefined silent-speech tasks and motion conditions, generating multiple repeated recordings for recognition-accuracy and motion-artifact analyses. So, this sample size was considered sufficient to demonstrate proof of concept, evaluate consistency of performance across participants, and characterize motion-artifact robustness under predefined conditions. Future work will include larger and more diverse participant cohorts. Ethics oversight All procedures involving human research participants were conducted in accordance with the experimental protocol approved by the Ethics Committee of Xiamen University (XDYX202501K02). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Information Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Information Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Reporting Summary Transparent Peer Review file Source data Source Data 1 Source Data 2 Source Data 1 Source Data 2