Surface EMG Dataset for Silent Speech Recognition
Year: 2026
Type: dataset
Abstract: Dataset OverviewThis dataset contains preprocessed surface electromyography (sEMG) recordings collected during silent articulation of 100 English words. The dataset was acquired for the development and evaluation of an EMG-based silent speech interface (SSI), with a particular focus on user-customizable vocabulary recognition using few-shot learning.The dataset contains recordings from 20 healthy participants (12 males and 8 females; mean age, 24.6 ± 2.28 years; age range, 20–29 years). None of the participants reported neurological or muscular impairments affecting facial movement or articulation.Each participant completed five recording sessions. During each session, participants silently articulated each of the 100 target English words once, resulting in 500 trials per participant.Only preprocessed sEMG signals are included in the publicly available files.Experimental ProtocolParticipants were seated comfortably in front of an LCD monitor and instructed to articulate each target word clearly without vocalization.The experiment consisted of five sessions, with each session containing all 100 target words presented in a randomized order. Before each trial, the target word and corresponding instructions were displayed on the screen. Participants initiated each trial by pressing the spacebar.After trial initiation, a fixation cross was presented for 0.5 s, followed by presentation of the target word. An auditory cue indicated the beginning of the articulation period, and participants were instructed to complete the silent articulation within 4 s. A second auditory cue indicated the end of the trial.Thus, the experimental structure for each participant was: 5 sessions × 100 words = 500 trials sEMG Data AcquisitionSurface EMG signals were recorded using a BioSemi ActiveTwo biosignal acquisition system (BioSemi B.V., Amsterdam, the Netherlands) at a sampling rate of 2,048 Hz.Six adhesive electrodes were positioned across the facial and submental regions to record muscle activity associated with silent articulation. Conductive gel was applied to improve electrode–skin contact and facilitate stable signal acquisition.A reference electrode was positioned on the nasal sidewall, and signals from the six recording electrodes were referenced to this electrode.Each trial consisted of 4 s of six-channel sEMG recording, corresponding to: 8,192 time points × 6 channels before preprocessing.Data Preprocessing A second-order Butterworth notch filter was applied to suppress power-line interference, followed by a second-order Butterworth band-pass filter from 20 to 210 Hz to remove low-frequency drift and retain the frequency range relevant to sEMG activity.The filtered sEMG signals were downsampled from 2,048 Hz to 512 Hz to reduce computational requirements.After preprocessing, each 4-s trial consisted of: 2,048 time points × 6 channels The files provided in this dataset contain these preprocessed sEMG signals.Directory StructureThe dataset is organized at the participant level: {Dataset_root}/ │ ├── S01.mat ├── S02.mat ├── S03.mat ├── ... └── S20.mat Each file corresponds to one participant, where: S{participant_index}.mat identifies the participant.For example: S1.mat contains all five recording sessions and all 100 target-word trials for participant 1.MATLAB File StructureWithin each participant .mat file, the preprocessed sEMG data are stored in the following MATLAB structure: high.x where high is a MATLAB structure and x contains the preprocessed sEMG data.The dimensions of high.x are: session × word × channel × time point Specifically: 5 × 100 × 6 × 2048 The dimensions correspond to:Dimension 1 – Session: five recording sessions (1–5)Dimension 2 – Word: 100 target-word indices (1–100)Dimension 3 – Channel: six sEMG recording channels (1–6)Dimension 4 – Time point: 2,048 samples corresponding to 4 s at 512 HzTherefore, for example: data = S.high.x(2, 15, :, :); represents the preprocessed six-channel sEMG recording obtained during session 2 for word index 15.For a single trial, the corresponding data can be interpreted as: 6 channels × 2048 time points after removing the singleton session and word dimensions.Word LabelsThe second dimension of high.x corresponds to the target-word index from 1 to 100.The word indices follow the 100-word vocabulary used in the associated experiment. Each index uniquely identifies one English target word and remains consistent across participants and recording sessions.Thus: high.x(session, word_index, channel, time_point) can be used to access the recording corresponding to a specific participant, session, and target word.Word_List.csv provides the mapping between the numerical word index used in the filenames and the corresponding target word.The file contains three columns in the following order: Index | Korean | English For example, word1.mat corresponds to the word listed under index 1 in Word_List.csv.The word index can therefore be used as the class label for word-level silent speech recognition.Associated StudyThe dataset was collected as part of the study:“A Metric-Based Few-Shot Learning Framework for User-Customizable Silent Speech Interfaces Using Electromyography.”The associated study investigated whether representations learned from a system-defined vocabulary could be generalized to previously unseen user-defined vocabulary using only a small number of labeled examples. The dataset was used to evaluate several metric-based few-shot learning approaches, including Siamese, Prototypical, Matching, and Relation networks.
Source: Figshare
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204223502