Deep sensing of breathing signal during conversational speech

In this paper, we show the first results on the estimation of breathing signal from conversational speech using deep learning algorithms. Respiratory diseases such as COPD, asthma, and respiratory infections are common in the elderly population and patients in health care monitoring and medical alert services in general. In this work, we compare algorithms for the estimation of a known respiratory target signal, measured by respiratory belt transducers positioned across the rib cage and abdomen, from conversational speech. We demonstrate the estimation of the respiratory signal from speech using convolutional and recurrent neural networks. The estimated breathing pattern gives respiratory rate, breathing capacity and thus might provide indications of the pathological condition of the speaker. Evaluation of our model on our database of breathing signal and speech yielded a sensitivity of 91.2 % for breath event detection and a mean absolute error of 1.01 breaths per minute for breathing rate estimation.