RESEARCH

HOME RESEARCH
Health Analytics
Other: Computation Methods for Health
Spoken Dialogs
Speech and Language
Learning Conditional Acoustic Latent Representation with Gender and Age Attributes for Automatic Pain Level Recognition
Abstract
Pain is an unpleasant internal sensation caused by bodily damages or physical illnesses with varied expressions conditioned on personal attributes. In this work, we propose an age-gender embedded latent acoustic representation learned using conditional maximum mean discrepancy variational autoencoder (MMD-CVAE). The learned MMD-CVAE embeds personal attributes information directly in the latent space. Our method achieves a 70.7% in extreme set classification (severe versus mild) and 47.7% in three-class recognition (severe, moderate, and mild) by using these MMD-CVAE encoded features on a large-scale real patients pain database. Our method improves a relative of 11.34% and 17.51% compared to using acoustic representation without age-gender conditioning in the extreme set and the three-class recognition respectively. Further analyses reveal under severe pain, females have higher maximum of jitter and lower harmonic energy ratio between F0, H1 and H2 compared to males, and the minimum value of jitter and shimmer are higher in the elderly compared to the non-elder group.
Figures
This is our overall framework. A conditional variational autoencoder architecture with maximum-mean-discrepancy criterion is used to learn acoustic representation for automatic pain classification.
This is our overall framework. A conditional variational autoencoder architecture with maximum-mean-discrepancy criterion is used to learn acoustic representation for automatic pain classification.
Keywords
pain | acoustic representation | age and gender | conditional variational autoencoder (CVAE)
Authors
Publication Date
2018/09/02
Conference
Interspeech 2018
Interspeech 2018
DOI
10.21437/Interspeech.2018-1298
Publisher
ISCA