Small Vocabulary Recognition Using Surface Electromyography in an Acoustically Harsh Environment

摘要

This paper presents results of electromyographicbased (EMG-based) speech recognition on a small vocabulary of 15 English words. The work was motivated in part by a desire to mitigate the effects of high acoustic noise on speech intelligibility in communication systems used by first responders. Both an off-line and a real-time system were constructed. Data were collected from a single male subject wearing a firefighter’s self-contained breathing apparatus. A single channel of EMG data was used, collected via surface sensors at a rate of 10 4 samples/s. The signal processing core consisted of an activity detector, a feature extractor, and a neural network classifier. In the off-line phase, 150 examples of each word were collected from the subject. Generalization testing, conducted using bootstrapping, produced an overall average correct classification rate on the 15 words of 74%, with a 95 % confidence interval of [71%, 77%]. Once the classifier was trained, the subject used the real-time system to communicate to a cellular phone and to control a robotic device. The realtime system was tested with the subject exposed to an ambient noise level of approximately 95 decibels.