Model ensemble for predicting heart and respiration rate from speech

Stress levels comprise a significant source of information in assessing human wellbeing including both mental and physical health. Interestingly, speech signals can be indicative of stress and may be used to infer related physiological markers, such as heart rate and respiration cycles. To this end, this work proposes a non-intrusive, low-cost, and automatic stress monitoring framework facilitating timely activation of stress relief methods and/or stress prevention. Initially, we design a multidomain speech feature extraction scheme able to reveal complementary stress-related characteristics. Subsequently, these are modeled by a synergistic framework able to encode both linear and non-linear relationships via suitably-learned support vectors and recurrent neural network. We employed an appropriate corpus encompassing recordings of job interviews, constructed based on a standardized experimental protocol. Importantly, such an approach outperformed the state of the art by 12.3% and 33.3% in predicting heart rate and respiration respectively.