MSEMG: Multi-Channel sEMG Signal Modeling

• MSEMG is a multi-channel surface electromyography framework that captures and processes muscle activity using non-invasive electrodes for applications like keystroke decoding and speech enhancement.
• It employs advanced acquisition methods, rigorous feature engineering, and state-space neural architectures to enhance signal fidelity and achieve robust denoising.
• Performance metrics such as SNR improvement and reduced error rates validate its effectiveness for real-time bioelectric interfacing and clinical signal processing.

MSEMG (Multi-channel Surface Electromyography): Modeling, Enhancement, and Denoising with State-Space Neural Architectures

Multi-channel surface electromyography (MSEMG) encompasses a set of acquisition, signal processing, and machine learning practices for capturing and modeling muscle activity via non-invasive electrodes. MSEMG methods are foundational for myoelectric human-computer interfaces, multimodal speech enhancement, and artifact-robust biomedical signal processing. Three research threads—keystroke transcription (Sivakumar et al., 2024), multimodal speech enhancement (Wang et al., 2022), and deep sEMG denoising (Liu et al., 2024)—elucidate contemporary approaches in data acquisition, feature modeling, neural architecture, and evaluation.

1. Acquisition Protocols and Hardware Implementation

State-of-the-art MSEMG studies employ multi-channel electrode arrays positioned to target compound muscle groups. In wrist-based keystroke decoding (Sivakumar et al., 2024), two wristbands each comprising 16 dry, gold-plated differential pairs (32 channels total), pre-filtered between 20–850 Hz and digitized at 2 kHz, are used to capture subtle neuromuscular signals during typing sessions. Facial EMG for multimodal speech enhancement (Wang et al., 2022) utilizes 35-channel arrays positioned on cheek and chin, sampled at 2048 Hz. In ECG artifact denoising (Liu et al., 2024), 12-channel upper-limb sEMG is recorded at 1 kHz.

Sampling protocols emphasize high temporal fidelity and consistent electrode-skin contact. Data transmission relies on Bluetooth or direct acquisition, with hardware-level bandpass filtering and high-precision ADCs (≥12 bit, ±6.6 mV input range). Annotations use automatic event loggers (e.g., keylogger synchronized to sEMG for ground-truth keystroke alignment). Subject pools range from 8 (speech) to 108 (typing) users, exceeding 5 million event-level labels in the largest datasets.

2. Signal Preprocessing and Feature Engineering

Raw MSEMG streams are processed through stringent filtering, normalization, and feature extraction regimes:

• Filtering and Alignment: High-pass filtering post-acquisition (e.g., 40 Hz, (Sivakumar et al., 2024)) is stacked atop analog band-pass filtering to attenuate drift and low-frequency motion artifacts. Sub-millisecond alignment (for multi-device systems) corrects for clock drift, essential in bimanual sensor configurations.
• Spectral Feature Extraction: Short-Time Fourier Transform (STFT) features with log-energy (33-bin log-frequency scale, 32 ms frames, 8 ms hop) dominate in keystroke and speech paradigms (Sivakumar et al., 2024, Wang et al., 2022). For EMG-only modalities, STFT features outperform raw/rectified signals; batch normalization across time-frequency stabilizes channel variance.
• Time-Domain Statistics: For facial EMG, TD15 features (mean, squared mean, rectified mean, squared highpass, zero-crossing rate) are computed per channel and window, followed by temporal stacking to form high-dimensional vectors (e.g., $\mathbb{R}^{5425}$ for 35 × 31 × 5, (Wang et al., 2022)).
• Augmentation: SpecAugment-style masking (temporal and frequency), electrode channel rotation, and temporal jitter augment MSEMG data to mitigate session variability, simulate hardware displacement, and enhance model robustness (Sivakumar et al., 2024).
• Superposition and SNR Control: In denoising contexts, ECG artifacts are superimposed onto sEMG at controlled SNRs, spanning −15 dB to 0 dB (Liu et al., 2024).

3. Neural Model Architectures for MSEMG

Recent MSEMG research converges on multi-stage neural architectures, differing in their treatment of temporal structure, sequence alignment, and multimodal fusion.

3.1 ASR-Inspired ConvNet-CTC for Decoding Keystrokes

The emg2qwerty baseline (Sivakumar et al., 2024) uses dual-band, rotation-invariant linear + ReLU subnets, followed by a concatenation and a 4-block time-depth-separable CNN (≈1 s receptive field). Output is mapped through a linear + CTC-softmax layer over the character set. The system employs Connectionist Temporal Classification (CTC) loss:

$$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$

A 6-gram character-level LLM (WikiText-103) is incorporated for improved keystroke sequence decoding via backspace-aware beam search.

3.2 Multimodal Fusion for Speech Enhancement

EMGSE (Wang et al., 2022) performs late fusion of 100D per-branch latent encodings using an FC → ReLU gate. Inputs are:

• Audio: 257D STFT log-magnitude features per frame.
• EMG: 5425D high-dimensional time-domain feature vectors.

Fusion output is processed by a 2-layer bidirectional LSTM (500 units per direction) and mapped back to a 257D enhanced spectrogram representation. Loss is L1-norm on log-magnitude spectrograms.

3.3 Mamba-based Denoising Network

The MSEMG model (Liu et al., 2024) integrates multi-scale convolutional Half-Normalized-Filter (HNF) encoder/decoder blocks with an intermediate Mamba selective state-space layer. The latter models sequence dependencies via a discretized SSM:

$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$

where $\overline{A} = \exp(\Delta A)$, $$\overline{B} = (\Delta A)^{-1}\left[\exp(\Delta A) - I\right]\Delta B$$, parameterized over a learned $\Delta$ (Liu et al., 2024). The Mamba block is followed by residual connections, and the network remains lightweight ($\sim2.8\times10^5$ params). MSE loss is used for optimization.

4. Performance Evaluation, Metrics, and Results

Quantitative assessment utilizes both direct signal metrics and task-specific outcome measures:

• Keystroke Decoding (CER): Character Error Rate $\mathrm{CER} = (S+D+I)/N$, where $S$, $D$, $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$0 are the numbers of substitutions, deletions, and insertions respectively and $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$1 is total ground-truth characters. Zero-shot generic models yield $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$251.8% CER, while light user personalization (10–20 min data) enables CER < 10% (Sivakumar et al., 2024).
• Speech Enhancement: Performance is assessed via PESQ (Perceptual Evaluation of Speech Quality), STOI (Short-Time Objective Intelligibility), and SI-SDR (optional). EMGSE outperforms audio-only systems, especially at low SNR (e.g., at −11 dB: PESQ gain $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$30.255, STOI gain $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$40.107). Cheek-only EMG achieves near-equal results to full-face arrays, enabling hardware simplification (Wang et al., 2022).
• Denoising: Metrics include SNR improvement, RMSE (signal), RMSE on average-rectified value (ARV), and mean-frequency (MF). On ECG artifact removal, MSEMG achieves $$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$5 dB SNR improvement, outperforming high-pass, template subtraction, FCN, and SDEMG models on all metrics (Liu et al., 2024).

Model	SNR$$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$6 (dB)	RMSE	Params
HP	13.885	$$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$7	N/A
TS	14.279	$$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$8	N/A
FCN	17.758	$$L_{\text{CTC}} = -\ln p_{\text{CTC}}(\mathbf{y}|\,\mathbf{X})$$9	$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$0
SDEMG	18.467	$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$1	$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$2
MSEMG	20.317	$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$3	$$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$4

5. Domain Shift, Personalization, and Robustness

Personalization and domain adaptation remain central for MSEMG performance:

• Zero-shot Transfer: Generic, non-personalized models display significant performance drop on new users or sessions (e.g., > 50% CER in keystroke prediction (Sivakumar et al., 2024)).
• Finetuning: Brief user-specific finetuning (< 30 min data) can restore performance (CER < 10%). This highlights the influence of cross-user anatomy, electrode placement, and behavioral variance.
• Augmentation and Invariance: Rotation channel augmentation and explicit rotation-invariant subnets attenuate session-to-session and placement-induced domain shift (Sivakumar et al., 2024). Robustness across SNRs and to unseen signal/patient pairs is empirically validated in MSEMG denoising (Liu et al., 2024).

6. Efficiency, Limitations, and Prospective Research

Recent architectural advances emphasize lightweight, real-time capable models without sacrificing signal fidelity:

• Efficiency: MSEMG denoiser is $$h_t = \overline{A} h_{t-1} + \overline{B} x_t, \quad y_t = C h_t$$51/5 the size of leading SDEMG diffusion models and is suitable for real-time deployment due to linear complexity with respect to sequence length (Liu et al., 2024).
• Feature Preservation: Highest SNR improvement and minimal error on key physiological features (ARV, MF) validate the value of joint CNN–state-space modeling for sEMG tasks.
• Limitations: Most models require supervised paired sEMG/label data, and generalization to new electrode montages or untrained domains remains challenging. Current approaches do not incorporate unsupervised or self-supervised adaptation for new muscle groups or hardware.
• Future Directions: Large-scale pretraining, self-supervised learning, spike-sorting, and dynamic attention-based fusion are noted as avenues to further improve generalization and adaptive robustness in MSEMG interfaces (Sivakumar et al., 2024, Wang et al., 2022).

7. Applications and Impact

MSEMG frameworks are integral to multiple next-generation bioelectric interfaces and biomedical processing tasks:

• Touch Typing via sEMG: Demonstrated feasibility for real-time, high-throughput text entry via wrist-worn sEMG with error rates approaching practical usability (Sivakumar et al., 2024).
• Speech Enhancement in Noise: EMG–acoustic fusion yields statistically significant intelligibility and quality improvements in adverse environments; practical deployment is particularly compelling for users in high-noise conditions or with speech impediments (Wang et al., 2022).
• Denoising for Prosthetic and Clinical sEMG: MSEMG’s denoiser achieves state-of-the-art performance, preserving signal detail in the presence of substantial cardiac interference without undue model complexity (Liu et al., 2024).

These results collectively underscore the maturation of MSEMG approaches for robust, high-bandwidth neural, muscular, and multimodal interface systems, with implications for assistive technology, neuroscientific inquiry, and future human-computer interaction modalities.