LowKeyEMG: Reduced-Key sEMG Typing

• LowKeyEMG is a reduced-key human-computer interface that maps 7 forearm gestures to text, merging 4 alphabetic and 3 control signals for efficient input.
• The system employs 64-channel sEMG acquisition and a dual-stage CNN pipeline with EEGNet to achieve over 90% accuracy in gesture detection and classification.
• Integrating a recurrent transformer (RWKV) for language disambiguation, LowKeyEMG delivers real-time typing speeds of up to 23.3 wpm with a 17% efficiency gain.

Searching arXiv for the LowKeyEMG paper and closely related EMG typing work. LowKeyEMG is a real-time human-computer interface for keyboardless text entry from surface electromyography (sEMG) that uses only 7 gesture classes decoded from the forearm rather than attempting full-alphabet EMG classification. Its central design choice is to reduce the English alphabet to 4 gesture keys and reserve 3 additional gestures for space and system interaction, then rely on the recurrent transformer-based LLM RWKV to resolve the resulting ambiguity in context. In the reported real-time experiments, participants achieved average one-handed keyboardless typing speeds of 23.3 words per minute, and gesture efficiency improved by 17% relative to typed phrase length; the paper further reports practical word-prediction accuracy under the 7-key constraint (Lee et al., 26 Jul 2025).

1. Reduced-key EMG typing paradigm

LowKeyEMG maps a small vocabulary of hand gestures onto 4 alphabetic signals plus 3 control signals, rather than attempting 26 or more distinct EMG classes. The alphabetic gestures are press thumb, press index finger, press ring finger, and press little finger. The control gestures are SPACE via ulnar-deviate wrist, DEL via lift little finger, and SEL via lift both middle and ring fingers (Lee et al., 26 Jul 2025).

Gesture type	Class or signal	Mapping
Alphabetic	1	press thumb
Alphabetic	2	press index finger
Alphabetic	4	press ring finger
Alphabetic	5	press little finger
Control	SPACE	ulnar-deviate wrist
Control	DEL	lift little finger
Control	SEL	lift both middle & ring fingers

The rationale stated for this reduced keyset is that decoding more than 10–12 EMG classes degrades rapidly in accuracy. By collapsing the 26 letters, plus punctuation, onto only 4 gesture classes, the front-end decoder achieves greater than 90% test accuracy on a same-day data set, while a recurrent LLM resolves ambiguity in context (Lee et al., 26 Jul 2025).

This design distinguishes LowKeyEMG from full-alphabet EMG typing systems. One earlier natural-typing system reconstructed typed text entirely from measured muscle potentials with 10-fold cross-validation mean character-level accuracy of 90.93%, but it operated on natural computer typing recorded from a physical keyboard and used a sequence model trained over continuous typing data from one typist (Crouch et al., 2021). LowKeyEMG instead targets one-handed, keyboardless text entry with an explicitly compressed symbol inventory. A plausible implication is that LowKeyEMG trades front-end class cardinality for stronger downstream language-model disambiguation.

2. sEMG acquisition and preprocessing

The system records sEMG at 2000 Hz using two custom saline-sponge electrode arrays wrapped around the right forearm, one placed proximally and the other distally. Each array comprises 32 electrodes, yielding 64 channels total, with a common reference and ground on the same limb. Data are average-referenced within each array and then processed by a 4th-order Butterworth bandpass filter from 10 to 999 Hz plus notch filters at 60, 120, 180, 240, and 300 Hz; prior to each session, the filtered signals are inspected to ensure that no line noise or movement artifacts remain (Lee et al., 26 Jul 2025).

Raw sEMG is segmented into non-overlapping 120 ms windows, corresponding to 240 samples at 2 kHz, for both gesture-detection and gesture-classification stages. Each window is represented as a $64 \times 240$ matrix of channels by time and is fed directly into a convolutional neural net, specifically EEGNet, without hand-crafted features. To improve robustness, training examples are time-shifted in 16 strides so that the activity peak moves across the window, and additive Gaussian noise with $\sigma = 0.25 \cdot \text{channel variance}$ is injected on the fly (Lee et al., 26 Jul 2025).

In methodological terms, this places LowKeyEMG in the end-to-end raw-signal tradition within EMG decoding. Related sEMG typing work also emphasized learned temporal features over hand-crafted descriptors and showed that high-accuracy character reconstruction is possible from raw differential waveforms (Crouch et al., 2021). LowKeyEMG’s hardware configuration differs in using 64 channels on one forearm rather than bilateral sleeves, and its preprocessing explicitly includes bandpass and notch filtering before neural decoding (Lee et al., 26 Jul 2025).

3. Neural decoding stack

LowKeyEMG uses a two-stage CNN pipeline in which both stages employ EEGNet with identical hyperparameters. The first stage performs gesture detection, asking whether there is an impending gesture in the 120 ms window. The second stage performs gesture classification over the 7 classes. The reported hyperparameters are temporal filters = 64, spatial filters = 4, dropout $p = 0.5$, average-pool factor = 2, optimizer = Adam, learning rate $1 \times 10^{-3}$, weight decay $1 \times 10^{-4}$, $\epsilon = 1 \times 10^{-3}$, max epochs = 30, and early stopping patience = 15 (Lee et al., 26 Jul 2025).

For both stages, the loss is cross-entropy,

$$L_{\rm CE} \;=\; -\sum_{i=1}^{C} y_i \,\log \hat y_i$$

with $C=2$ for detection and $C=7$ for classification (Lee et al., 26 Jul 2025).

Real-time decoding operates on windows that slide every 120 ms for gesture detection. Once a gesture is detected, the window is realigned to the sample of maximum average amplitude and re-classified. A minimum inter-gesture interval of 0.2–0.4 s is imposed to prevent spuriously rapid detections (Lee et al., 26 Jul 2025).

This pipeline reflects a segmentation-first formulation rather than continuous-sequence transcription. That differs from Connectionist Temporal Classification-based continuous typing recognition, where the network outputs frame-wise probabilities over a character inventory at fixed temporal resolution and learns alignments implicitly (Crouch et al., 2021). LowKeyEMG’s two-stage detector-classifier division is consistent with its small gesture vocabulary and explicit real-time interaction loop.

4. Language-model disambiguation and real-time inference

Text decoding is performed with the pretrained RWKV “recurrent transformer,” specifically a 169 M-parameter Pile checkpoint. The paper emphasizes two properties: recurrent update means that next-token inference cost is $\mathcal O(1)$ per token, independent of context length, and the maximum context length is much greater than a 4-gram LLM, offering richer disambiguation of 4-key input (Lee et al., 26 Jul 2025).

After each classified gesture $\sigma = 0.25 \cdot \text{channel variance}$0, a beam search with width $\sigma = 0.25 \cdot \text{channel variance}$1 merges a language-model score $\sigma = 0.25 \cdot \text{channel variance}$2 with a character-model penalty for Levenshtein-distance-1 candidates. On an Intel i7-12800H plus NVIDIA RTX 3070 Ti Mobile, each beam-search step takes

$\sigma = 0.25 \cdot \text{channel variance}$3

on average, which the paper reports as real-time decoding well under perceptual thresholds (Lee et al., 26 Jul 2025).

The distinction between the reduced-key front end and the language-model back end is central. The 7 gestures do not directly encode words or a full alphabet; rather, the LLM resolves ambiguity introduced by mapping the alphabet onto only 4 alphabetic gesture classes. This reduces the burden on EMG classification while increasing the importance of contextual text priors.

A broader pattern appears in adjacent EMG-to-text research. In unvoiced speech EMG, a frozen LLM paired with a lightweight EMG adaptor achieved an average word error rate of 0.49 on a closed-vocabulary task, indicating that strong text priors can compensate for low-bandwidth or ambiguous biosignal inputs (Mohapatra et al., 30 May 2025). This suggests that LowKeyEMG’s use of RWKV is not incidental but part of a wider shift toward language-model-assisted decoding of constrained EMG channels.

5. Experimental protocol and quantitative performance

The reported participant cohort consists of $\sigma = 0.25 \cdot \text{channel variance}$4 healthy right-handed adults aged 25–33, each of whom gave IRB-approved informed consent and received $20/h compensation. Two task families were used. The rhythm task collected 560 prompted gestures per day, 80 per class, under 6 force and pose variants to train, validate, and test the CNNs. The typing task used 42 short passages of 4 or more sentences; the first 40 characters of the last sentence served as prompt, and preceding context may or may not be provided (Lee et al., 26 Jul 2025).

Three typing conditions were evaluated: Base with no context, no completion, and no distance-1 matching; Completion only with no context but exact, distance-1, and prefix matches allowed; and Completion + context with the same matching plus up to 3-sentence prompt context. The evaluation metrics are words per minute, gestures per character and efficiency gain, and top-$\sigma = 0.25 \cdot \text{channel variance}$5 candidate accuracy, especially $\sigma = 0.25 \cdot \text{channel variance}$6 (Lee et al., 26 Jul 2025).

The reported interleaved average speeds increase progressively from Base to Completion to Completion + Context: 14.4, 17.2, and 19.5 words per minute, respectively. In isolated expert runs under Completion + context repeated three times, the individual results are H1 27.8 wpm, H2 16.4 wpm, and H3 25.6 wpm, for a mean of 23.3 wpm (Lee et al., 26 Jul 2025).

Gesture efficiency, measured as error-corrected gestures per character, drops from 1.11 to 0.83, corresponding to a 17% relative efficiency gain with the full system. For simulated full passages, top-3 word-prediction accuracy is reported as 99.2% with RWKV, versus 85.8% using a 4-gram word model on the same 7-key mapping; top-1 accuracy with 4 keys is 96.9% versus 66.6% for the 4-gram model, and top-2 and top-3 accuracies are 98.9% and 99.2% (Lee et al., 26 Jul 2025).

The abstract reports a closely related but not identical summary figure: when typing with only 7 keys, LowKeyEMG can achieve 98.2% top-3 word accuracy (Lee et al., 26 Jul 2025). The coexistence of 98.2% in the abstract and 99.2% in the detailed results indicates that the paper presents more than one top-3 figure under slightly different evaluation descriptions. The concrete result is therefore that top-3 performance is reported as very high under the reduced-key paradigm, with the detailed passage-level simulation favoring RWKV over a 4-gram baseline (Lee et al., 26 Jul 2025).

6. Relation to prior EMG text interfaces, applications, and limitations

LowKeyEMG is situated against prior EMG typing work that attempted full-QWERTY or large-vocabulary decoding. The paper states that comparison to prior EMG typing methods shows that LowKeyEMG achieves near-state-of-the-art speeds and accuracy even though it uses only 7 gestures, whereas prior work suffered high character error rate when decoding more than 10 gestures (Lee et al., 26 Jul 2025). A directly relevant contrast is the natural typing recognition system that decoded 32 characters plus separator and no-output labels from continuous sEMG, obtaining 90.93% mean character-level accuracy but relying on natural keyboard typing, continuous-sequence modeling, and a single typist dataset (Crouch et al., 2021).

Another relevant comparison comes from EMG keylogging as a side-channel task. A between-subject attack using the Myo Armband reported mean balanced accuracy of about 76% for keystroke detection and mean top-3 key accuracy of about 32% on 52 classes, with EMG identified as the most prominent source of information compared to accelerometer and gyroscope streams (Gazzari et al., 2021). That result underscores the difficulty of fine-grained multiclass keystroke identification from forearm EMG, and it is therefore consistent with LowKeyEMG’s decision to reduce the symbol inventory before language-model disambiguation.

The applications listed for LowKeyEMG are assistive communication for users with motor impairments such as ALS, stroke, and SCI, low-dimensional input modalities such as eye-tracking, switch devices, and Morse-like binary coding, and on-device low-latency text entry in VR/AR or wearable contexts where full keyboards are infeasible (Lee et al., 26 Jul 2025). These are application domains where only a few reliable EMG classes may be extractable or where input bandwidth is inherently constrained.

The reported evidence should nevertheless be read in light of the stated protocol: same-day front-end testing, a small cohort of three healthy right-handed adults, and a typing task structured around prompted passages and varying contextual assistance (Lee et al., 26 Jul 2025). This suggests that the principal contribution is a system design principle—collapse the alphabet onto a handful of robustly decoded gestures and offload disambiguation to a lightweight recurrent transformer—rather than a claim that full open-ended typing has been solved under all EMG conditions. Within that scope, the paper shows that one-handed, keyboardless typing at approximately 20 words per minute or higher is achievable with very high top-3 candidate accuracy under a 7-key EMG interface (Lee et al., 26 Jul 2025).