Authors:
Michael Wand
;
Christopher Schulte
;
Matthias Janke
and
Tanja Schultz
Affiliation:
Karlsruhe Institute of Technology, Germany
Keyword(s):
EMG, EMG-based Speech Recognition, Silent Speech Interface, Electrode Array.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Data Manipulation
;
Devices
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Human-Computer Interaction
;
Methodologies and Methods
;
Neurocomputing
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition
;
Physiological Computing Systems
;
Physiological Processes and Bio-Signal Modeling, Non-Linear Dynamics
;
Sensor Networks
;
Soft Computing
;
Speech Recognition
;
Wearable Sensors and Systems
Abstract:
An electromygraphic (EMG) Silent Speech Interface is a system which recognizes speech by capturing the electric potentials of the human articulatory muscles, thus enabling the user to communicate silently. This study is concerned with introducing an EMG recording system based on multi-channel electrode arrays. We first present our new system and introduce a method to deal with undertraining effects which emerge due to the high dimensionality of our EMG features. Second, we show that Independent Component Analysis improves the classification accuracy of the EMG array-based recognizer by up to 22.9% relative, which is a first example of an EMG signal processing method which is specifically enabled by our new array-based system. We evaluate our system on recordings of audible speech; achieving an optimal average word error rate of 10.9% with a training set of less than 10 minutes on a vocabulary of 108 words.