Online Detection of Speaking from Respiratory Measurement Collected in the Natural Environment

We present a novel method to detect when a person is speaking using respiratory measurements collected in the natural enviroment. A speaker’s respiration pattern is sampled from a respiratory inductive plethysmograph (RIP) band worn around the speaker’s chest. Ratio of inhalation duration to exhalation duration (IE ratio) has traditionally been used to detect speaking in controlled lab environment [7]. But we find that IE ratio is inadequate in the natural environment. We propose several new features to be used along with IE ratio. Using various statistics over these features in a decision tree, we are able to obtain >93% accuracy in classifying respiration measurements into “speaking” or “silence” states with 10-fold cross validation. Our demonstration will show real-time capture of the respiration signal, computation of features, and detection of speaking and silence all on a mobile smartphone.