Ultrasensitive textile strain sensors redefine wearable silent speech interfaces with high machine learning efficiency

Silent speech interfaces (SSI) have emerged as a solution for communication when spoken language is hindered by noisy environments or physiological conditions (e.g., stroke, cerebral palsy, Parkinson's disease, post-laryngeal surgery). By analyzing nonvocal human signals, SSIs decode speech in silent conditions. A key challenge is developing a wearable system that is comfortable, durable, precise, and efficient across users and scenarios. Recent efforts target both sensing devices and algorithmic models. Speech-related neural impulses originate in the central nervous system, travel to the vocal cords, and are articulated via facial movements to produce sounds. In wearable SSI, capturing throat micromovements with sufficient sensitivity can enable accurate decoding while preserving comfort and practicality. In pursuit

The work situates itself among SSI approaches using electromyography (EMG), mechanical/strain sensors, and other modalities. In real-world SSIs, key noise sources include sensor flicker noise, environmental sound, and physiological artifacts (breathing, swallowing, neck motion). Prior SSI pipelines often convert 1D time-series signals into 2D representations (e.g., via Fourier transforms) for 2D neural networks, especially when using multi-channel arrays to exploit spatial resolution or when single-channel sensors provide low information density. However, 2D methods increase computational complexity and are less suitable for energy-constrained wearables. Textile-based strain sensors made via printing/coating have typically exhibited relatively low gauge factors within small strain ranges, limiting their ability to capture rich information for distinguishing words. The present study addresses these constraints by designing a highly sensitive textile strain sensor with ordered cracks to increase information density and enable efficient 1D neural processing, contrasting with prior lower-sensitivity single-channel systems that relied on heavier 2D feature extraction.

Materials and ink preparation: TIMREX KS 25 graphite (25 µm) was used with sodium deoxycholate (SDC, ≥97%) as surfactant and sodium carboxymethyl cellulose (CMC-Na, average MW 700,000) as binder. DI water was the solvent. SDC was dissolved in DI water at 5 g/L. Graphite flakes were added at 100 g/L and mixed (500 rpm, 30 min). Few-layer graphene flakes were produced by high-pressure homogenization (HPH, PSI-40) using a dual-slot chamber (D200D, 200 µm) at 700 bar for 70 exfoliation cycles. CMC-Na (10 g/L) was added to the dispersion to stabilize flakes and adjust viscosity, stirring 3 h at room temperature. Sensor fabrication with ordered cracks: Textile substrate (95% bamboo fibers, 5% elastane) was UV-ozone treated (30 min) to improve hydrophilicity and adhesion. Graphene ink (100 g/L) was screen-printed (mesh 90T/230 mesh per inch) into rectangular patterns. To form a continuous top graphene layer and enable ordered cracks, 7 printing passes were performed, each with a 2 mL ink dose, drying at room temperature with N2 blow for 1 min between prints. Post-print anneal at 80 °C for 5 min. The printed substrate was prestretched by 5% strain to induce ordered through-cracks aligned with the woven textile’s stress concentrations. Characterization: AFM (Bruker Icon) was used to measure lateral size, thickness, and aspect ratios of graphene flakes (100 flakes across three ~20 µm × 20 µm scans). SEM (Magellan 400) characterized textile morphology and crack structures. Electromechanical measurements used a Deben Microtest 200 N tensile stage/INSTRON system and a Keithley 2400 SMU. Resistance responses under repetitive strains were recorded to evaluate linearity, hysteresis, detection limit, frequency response, and durability. Data acquisition: The strain sensor was integrated into a smart choker; copper tape contacts were affixed 1 cm apart. A potentiostat (EmStat4S, PalmSens) supplied 1 V and measured current. Sampling at 500 Hz; each word sample lasted 3 s. Data collection reflected real-world variability: participants wore the choker comfortably with non-rigid positioning/tightness. Noise augmentation and model: To enhance robustness without filtering, a random noise window augmentation was used. Noise-only segments (breathing, head turns) were recorded while wearing the choker; multiple same-length noise windows were randomly sampled and overlaid onto speech samples to create augmented data (each original sample yielded four augmented samples). A lightweight 1D end-to-end CNN with residual blocks was designed: initial Conv1D with 64 filters (kernel size 7) + BatchNorm + ReLU; residual blocks of paired Conv1D (kernel size 3) with BatchNorm and ReLU; dropout 0.2; max-pooling for downsampling; fully connected layers leading to classification. Inputs were 3 s signals (1500 points at 500 Hz). Implementation used Python 3.8.13, Miniconda 3, PyTorch 2.0.1 with Apple MPS acceleration. Datasets and evaluation: Three datasets from three participants: (1) 20 high-frequency English words (10 verbs + 10 nouns), (2) 10 easily confusable words differing by one phonetic element (e.g., book/look, sheep/ship, record variants), (3) five long words read at varying speeds. For each class: 100 samples (80 train, 20 test). Transfer learning experiments used the baseline model (trained on Dataset 1) as a pretrained model, fine-tuned with limited samples from three new users and ten new words.

- Sensor performance: Ordered through-crack graphene-coated textile achieved an ultrahigh gauge factor (GF) of 317 within ≤5% strain, a 420% improvement over printed/coated textile strain sensors reported to date. The sensor displayed linear relative resistance response with low hysteresis within small strains (≤5%). It resisted tensile frequency interference, aiding recognition of words spoken at different pitches. Detection limit was 0.05%. Durability exceeded 10,000 stretch–release cycles at 1.5% strain with stable performance. The system was unresponsive to 100 dB environmental sound noise (100% unresponsive in tests), yet highly sensitive to throat micromovements. The device also exhibited strong adhesion and stability of ordered cracks, minimizing drift. - Model efficiency and accuracy: A lightweight 1D CNN leveraged the high information density from the sensor to reduce computational load by 90% while maintaining high accuracy. On Dataset 1 (20 common words), accuracy was 95.25%. On Dataset 2 (10 confusable words), accuracy was 93%. On Dataset 3 (five long words at varying speeds), accuracy was 96%. FLOPS and channel usage were significantly lower than state-of-the-art SSI models, indicating time and energy efficiency suitable for wearables. - Generalization: Transfer learning from the baseline model enabled rapid adaptation. With only 15–20 samples per class, accuracy reached ~80% on new users and new words (improvements of 53% and 43% over training from scratch, respectively). With 30 samples per class, accuracy reached ~90% for both new users and words. t-SNE visualizations showed improved separability after fine-tuning, including better discrimination of confusable pairs (e.g., book vs. look). - Robustness: R-CAM visualizations indicated the model focused on key micromovement regions rather than noise. The system was robust to DC offset variations due to different choker tightness/placement and to physiological noise, aided by noise-window augmentation.

The study addresses the central SSI challenge of combining comfort, durability, precision, and computational efficiency in a wearable system. By engineering ordered through cracks in a continuous graphene layer printed on a textile matrix, the sensor achieves ultrahigh sensitivity within the small strain regime relevant to throat micromovements. This high sensitivity yields information-dense single-channel signals, enabling an efficient 1D end-to-end neural network rather than computationally heavy 2D methods. Together, the sensor and model deliver high decoding accuracy with substantially reduced FLOPS, supporting edge deployment in wearable devices. The system demonstrates robustness to major real-world noise sources (flicker, physiological artifacts) and environmental sound (unresponsive to 100 dB), as well as resilience to variations in wearing tightness and position. Transfer learning results show rapid adaptation to new users and vocabularies with limited data, suggesting practical paths for personalization and scalability. Overall, the synergy of materials/process optimization and tailored ML design directly addresses SSI requirements for real-world, energy-efficient, and accurate silent communication.

This work demonstrates a biocompatible, scalable textile-integrated graphene strain sensor with ordered through cracks, achieving a gauge factor of 317 (≤5% strain), 0.05% detection limit, and >10,000-cycle durability, enabling precise capture of throat micromovements. Leveraging the sensor’s high information density, a lightweight 1D CNN achieves up to 95.25% accuracy on common words while reducing computational load by 90%, with strong robustness to noise and wearing variations. The system decodes a wide vocabulary, adapts quickly to new users and words via transfer learning, and sets a benchmark for energy-efficient SSI suitable for wearable deployment. As model accuracy continued to improve with more training samples without saturation, expanded datasets and broader vocabularies/users present avenues for further performance gains and real-world integration.

Evaluation involved three primary participants and specific English vocabularies (20 common words, 10 confusable pairs, and five long words), with generalization tested on three new users and ten new words. While transfer learning showed strong adaptation with limited data, performance was assessed on relatively small datasets; results indicated accuracy continued to improve with more training samples.