Multimodal sEMG and Accelerometer Dataset for Silent Speech Recognition
Year: 2026
Type: dataset
Abstract: Dataset OverviewThis dataset contains surface electromyography (sEMG) and inertial measurement unit (IMU) recordings collected during silent articulation of 100 English words. The dataset was acquired for the development and evaluation of a multimodal silent speech interface (SSI) that combines electrical muscle activity with articulatory motion information.The publicly available dataset contains recordings from 10 healthy participants (6 males and 4 females; 24.6 ± 2.50 years). Each participant completed five recording sessions. During each session, participants silently articulated 100 words commonly used in daily life. The target words were presented in a randomized order, and each articulation was recorded for approximately 4 s.The dataset contains both sEMG and IMU recordings for each participant, session, and target word. The correspondence between word indices and target words is provided separately in Word_List.csv.Data AcquisitionsEMGSurface EMG signals were recorded from six electrodes positioned over facial and neck regions associated with speech-related muscle activity. Signals were acquired using a BioSemi ActiveTwo system (BioSemi B.V., Amsterdam, Netherlands) at a sampling rate of 2,048 Hz.Both raw and preprocessed sEMG signals are provided.The preprocessing applied to the sEMG signals included:downsampling from 2,048 Hz to 512 Hz;60-Hz notch filtering to suppress power-line interference; and20–210 Hz band-pass filtering.Each trial corresponds to a 4-s silent articulation.IMUArticulatory motion was recorded using five MPU-9250 sensor units positioned on the facial and submental regions. The sensors were connected through an I2C multiplexer to an ESP32-S3 microcontroller, and signals were acquired at approximately 100 Hz.Each IMU recording contains data from five sensors and six measurement components:x, y, z, roll, pitch, and yaw.The stored data have the following dimensional organization: sensor × measurement component × time point where:sensor = 5;measurement component = 6 (x, y, z, roll, pitch, and yaw); andtime point ≈ 400 samples for a 4-s recording at approximately 100 Hz.Because the actual number of acquired IMU samples could vary slightly across trials, some recordings contain more or fewer than exactly 400 time points. In the associated study, recordings containing more than 400 samples were truncated to the first 400 samples. Recordings containing slightly fewer than 400 samples were interpolated to obtain a standardized length of 400 time points.Only the x, y, and z acceleration components were used in the associated silent speech recognition study. After standardization to 400 time points, these signals were further downsampled from 100 Hz to 50 Hz for model input.The IMU files provided in this dataset contain the recordings prior to these preprocessing procedures.Experimental ProtocolEach participant completed five recording sessions. In each session, the participant silently articulated each of 100 target words once.The target words were presented in a randomized order using a custom graphical user interface. Participants initiated each trial and were instructed to complete the silent articulation within the 4-s recording period. A retry function was available during data acquisition when an articulation error occurred.The dataset is therefore organized according to three principal levels:participant → session → wordDirectory StructureThe dataset is organized as follows: {Dataset_root}/ │ ├── Word_List.csv │ ├── S01/ │ ├── EMG/ │ │ ├── session_1/ │ │ │ ├── raw_word1.mat │ │ │ ├── word1.mat │ │ │ ├── raw_word2.mat │ │ │ ├── word2.mat │ │ │ └── ... │ │ ├── session_2/ │ │ ├── session_3/ │ │ ├── session_4/ │ │ └── session_5/ │ │ │ └── IMU/ │ ├── session_1/ │ │ ├── word1.mat │ │ ├── word2.mat │ │ └── ... │ ├── session_2/ │ ├── session_3/ │ ├── session_4/ │ └── session_5/ │ ├── S02/ │ └── ... │ └── S10/ └── ... S{participant_index} identifies the participant, session_{session_index} identifies one of the five recording sessions, and word{word_index} identifies the target word.EMG File DescriptionTwo types of sEMG files are provided for each trial.raw_word{word_index}.matThis file contains the original sEMG recording.MATLAB variable:save_data_rawData dimensions: time point × channel Sampling rate: 2,048 HzRecording duration: 4 sNumber of time points: 8,192Number of channels: 6Thus, each file contains an 8192 × 6 sEMG data matrix.word{word_index}.matThis file contains the preprocessed sEMG signal.MATLAB variable:save_dataData dimensions: time point × channel Sampling rate: 512 HzRecording duration: 4 sNumber of time points: 2,048Number of channels: 6Thus, each file contains a 2048 × 6 preprocessed sEMG data matrix.IMU File Descriptionword{word_index}.matThis file contains the IMU recording corresponding to one target-word trial.MATLAB variable:save_dataThe data are organized as: sensor × measurement component × time point with: 5 sensors × 6 components × approximately 400 time points The six measurement components correspond, in order, to: x, y, z, roll, pitch, yaw The exact number of time points may differ slightly among recordings because of variations during acquisition. The files contain the original recordings without standardization of the time dimension.For the analyses reported in the associated study, only the x, y, and z components were used. Recordings longer than 400 samples were truncated to the first 400 samples, whereas recordings slightly shorter than 400 samples were interpolated to obtain 400 time points. The resulting signals were subsequently downsampled to 50 Hz for model input.Associated StudyThe dataset was collected as part of the study:“Hybrid Silent Speech Interface Integrating Surface Electromyography and Accelerometer for Enhanced Accuracy in Clinical Applications.”The associated study investigated a hybrid silent speech interface that integrates sEMG-derived muscle activity with articulatory motion information for silent word recognition.Although the associated study included both healthy participants and individuals who had undergone laryngectomy, the present publicly available dataset contains only the recordings from the 10 healthy participants.
Source: Figshare
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204266366