Published October 12, 2022 | Version 1.0.0
Dataset Open

Non-acoustic speech sensing system based on flexible piezoelectric

  • 1. Beihang University
  • 2. Tsinghua University


The non-acoustic speech sensing system based on flexible piezoelectric is designed to satisfy specific needs around testing device models (in high-noise, complex environments). The system collected vibration signals from the jaws of six males and five females containing ten different control commands at 90 dB of background noise. The dataset is reliable with high intelligibility and is able to achieve 93.7% recognition accuracy by calculation. In general, this paper provides a non-acoustic speech dataset for Mandarin, including the parts collected, the number of people collected, and the environment.

The dataset is available at:

The data descriptor paper with details of data collection and cleaning process is under submission. For proper citation of the manuscript, please refer to the latest version of this dataset which includes the details.

This dataset and its descriptor paper were created by:

Shiji Yuan, Ying Sun, Shuai Wang, Xinlei Chen,Ying Ding,Dezhi Zheng , Shangchun Fan

For questions or suggestions, please e-mail Shuai Wang <>

Ten common words were chosen as the core of the vocabulary in this dataset. These ten command words can be used for commands in IoT or robotics applications: "forward", "backward", "right", "left", "stop", "up", "down", "draw", "drop", and "reset".

The recording was carried on by software named Adobe Audition2022. We set monophonic recording, 16-bit storage format, and 16 kHz sampling frequency before recording and saved the recorded voice in wav format.  The dataset is provided with two storage rules, which are stored by subject number and command number as classification. In the first rule, the speech data of 11 subjects were stored in different folders with the subject serial number as the folder name. Each folder contains subfolders categorized by command. In the second rule, the speech data of ten commands are stored in different folders, and the names of the folders are the command contents. The subject number, command number and record order are given for each data entry. For example, the data obtained when subject 1 recorded command 10 for the first time was labeled as "1-10_1".

After the data collection process, a filtering algorithm for automatic detection of low non-acoustic speech data was designed to remove problematic data that were very short or very quiet.The script of the data filtering algorithm is provided in this repository.  

For specific detail of the data filtering process, please refer to the script (speech data filtering algorithm in MATLAB) in this repository and the data descriptor paper.

The dataset in this repository is the processed version. The raw dataset and removed audio files are not included in this repository.

File list:

Non-acoustic Speech

speech data filtering




Non-acoustic Speech

Files (695.6 MB)

Name Size Download all
695.6 MB Preview Download
4.0 kB Preview Download
29.2 kB Preview Download