SilentWear: An Ultra-Low Power Wearable Interface for EMG-Based Silent Speech Recognition

This repository provides a multi-session surface electromyography (EMG) dataset for vocalized and silent speech recognition, recorded using a wearable neckband interface.

The dataset is designed to support research in:

• EMG-based speech decoding
• Human–machine interaction (HMI)
• Assistive communication technologies
• Ultra-low-power wearable AI systems

The data were collected using SilentWear, an unobtrusive, ultra-low-power EMG neckband designed for silent and vocalized speech detection.

---

Dataset Description

The dataset includes recordings from:

• 4 subjects (3 male, 1 female)
• Vocalized and silent speech conditions
• 8 HMI commands:
  up, down, left, right, start, stop, forward, backward
  plus a rest (no-speech) class
• 3 recording days per subject
• Multiple sessions, collected over 3 days, each containing:
  • 5 vocalized batches.
  • 5 silent batches
• Each batch contains 20 repetitions of each word, plus rest.

This structure enables evaluation under multi-day conditions, supporting research on robustness to electrode repositioning and inter-session variability.

Further details on the data collection methodology are available at:
https://arxiv.org/placeholder

---

Repository Organization

The repository contains two subfolders:

1️⃣ data_raw_and_filt

This folder contains full-length EMG recordings for each subject, condition, session, and batch.

Each file: - Contains raw EMG signals - Contains filtered EMG signals (4th-order high-pass at 20 Hz + 50 Hz notch) - Is stored in .h5 format
- Uses the HDF5 key "emg"

Directory structure example:

data_raw_and_filt/
└── S01/s
    └── silent/
        └── sess_1_batch_1.h5
        .
        .
        └── sess_3_batch_5.h5
    └── vocalized/
        └── sess_1_batch_1.h5
        .
        .
        └── sess_3_batch_5.h5
└── S02
└── S03
└── S04

---

Example: Loading a File

import pandas as pd

df = pd.read_hdf("data_raw_and_filt/S01/silent/sess_1_batch_1.h5", key="emg")
df.head()

---

File Content Structure (data_raw_and_filt)

Each .h5 file contains:

Group	Columns	Description
Raw EMG	Ch_0–Ch_15	Raw sEMG samples
Filtered EMG	Ch_0_filt–Ch_15_filt	High-pass (20 Hz) + 50 Hz notch
Labels	Label_int, Label_str	Integer and string class labels
Metadata	session_id, batch_id	Session and batch identifiers

2️⃣ wins_and_features

• Non-overlapping windowed segments
• Raw and filtered signals
• Extracted time-frequency features

These files can be directly used for model training or benchmarking.

Code and Usage

The dataset is designed to be used in conjunction with the SilentWear repository:

https://github.com/pulp-bio/silent_wear

Please refer to the repository README.md for:

• Data loading utilities
• Preprocessing pipelines
• Training scripts
• Evaluation scripts

The repository creates the files contained in wins_and_features folder; these files are then used for model training.

Alternatively, you may directly use the data_raw_and_filt folder to:

• Build custom dataloaders
• Train your own architectures
• Benchmark novel EMG decoding methods

---

Contributing

We aim to promote standardized evaluation and fair comparison across models.

We strongly encourage contributions of trained models and evaluation results to:

https://github.com/pulp-bio/silent_wear

Please refer to the repository README for submission guidelines.

---

Citation

If you use this dataset, please cite:

@online{spacone_silentwear_26,
  author = {Spacone, Giusy and Frey, Sebastian and Pollo, Giovanni and Burrello, Alessio and Pagliari, J. Daniele and Kartsch, Victor and Cossettini, Andrea and Benini, Luca},
  title = {SilentWear: An Ultra-Low Power Wearable Interface for EMG-Based Silent Speech Recognition},
  year = {2026},
  url = {coming soon}
}

📄 License

See the LICENSE file for the full license text.

This project makes use of the following licenses:

• Apache License 2.0 — See the LICENSE file for the full license text.

• Images are under the the Creative Commons Attribution 4.0 International License - see the LICENSE.images file for details.

👨‍💻 Contributors

Silent-Wear has been developed at ETH Zürich, by the PULP-Bio:

• Giusy Spacone (Conceptualization, Experimental Design, Development)

• Sebastian Frey (PCB design, Firmware, Documentation)

• Fiona Meier (Hardware Development)

• Giovanni Pollo (Experimental Desing, Data Collection, Documentation)

• Prof. Luca Benini(Supervision, Conceptualization)

• Dr. Andrea Cossettini(Supervision, Project administration)