Predicting Keystrokes from Electromyography Signals

Peike Li, Wenxuan Karen Li, Debajyoti Chakrabarti, Qizhao Chen

ECE C247A Course Research Project · Winter 2026

EMG Sequence Modeling LSTM Transformers CTC PyTorch

Report Code Reference Paper ← Back to Home

Overview

Surface electromyography (sEMG) contains muscle-activity signals that can be used to infer typed keystrokes without a physical keyboard. We studied neural sequence models for decoding keystroke sequences from multichannel EMG recordings, focusing on architectures that can capture temporal dependencies in noisy high-dimensional signals.

We compared a CNN baseline with LSTM and Transformer-based models and then studied how performance changes with training-data availability, electrode density, data augmentation, and temporal sampling rate.

Sequence Models for EMG Decoding

LSTM:

Processes multichannel EMG sequences temporally
MLP feature projection
Stacked LSTM layers
CTC loss for sequence alignment

Transformer:

Positional encoding
Self-attention over temporal features
CTC output
Also tested a temporal-convolution (TDSConv) + Transformer variant

Model Performance

Bar chart comparing test character error rate: 23.21% for the CNN baseline versus 12.06% for the LSTM model. — Test character error rate across evaluated model architectures. Lower is better.

Method	Test CER (%)
Baseline (CNN)	23.21
LSTM (RNN)	12.06
Transformer (no downsampling)	38.08
Transformer-only	100.00
TDSConv + Transformer	30.69

Under this dataset and training setup, the LSTM generalized best. This does not imply that Transformer architectures are generally worse for EMG decoding — only that, among the configurations tested here, the LSTM achieved the lowest test CER.

What Affected Decoding Performance?

Four-panel figure: model comparison, test CER vs. training-data size, vs. number of channels, and vs. sampling rate. — Evaluation of the LSTM-based model under different experimental settings.

Training data: performance degrades significantly as the available training set becomes smaller
Electrode density: reducing from 32 to 16 channels causes only a modest degradation, while reducing to 8 channels degrades performance more strongly
Sampling frequency: an intermediate sampling rate performed best in the tested configuration
Augmentation: moderate augmentation improved generalization, while excessive augmentation degraded performance

Training Behavior

Training and validation CER curves across the evaluated architectures. — Training and validation CER across the evaluated architectures.

The LSTM showed the strongest validation behavior and the smallest generalization gap among the tested sequence models.

Key Takeaway

Temporal structure mattered more than model complexity: under this dataset and training setup, the LSTM substantially outperformed the CNN baseline and the tested Transformer variants.

LSTM reduced CER from 23.21% to 12.06%
CTC enabled sequence-level keystroke decoding
Performance depended strongly on training-data availability
Moderate sensor reduction remained viable
Sampling rate and augmentation had non-trivial effects on generalization

Report Code Reference Paper ← Back to Home