|
Introduction
|
Motivation Voicemail has become a crucial part of our personal and professional lives. The number of messages that accumulate in our mailboxes force us to consider new ways of prioritizing them. Currently, we are forced to actively listen to all messages in order to find out which ones are important and which ones can be attended to later on. A system that can detect the significant emotions in a message can change the way we perceive voicemail: 1) Users can be alerted of new messages along various affective axes (“You have an urgent message!” or “This messages sounds a little sad!”). This would partially shift the burden of prioritization from the user to the system. 2) Users can start searching amongst voicemail messages using affect (“Let me hear all the happy messages i received today”) 3) We can start to build speaker dependent emotion profiles for those who call frequently and monitor variations from their average emotional behavior. This project investigates application 1 : speaker independent emotion recognition in voicemail messages for smarter voicemail alerts. Approach
Our approach is based on extracting salient acoustic features from various voicemail messages and training a number of Hidden Markov Models . We have decided to look at four axes of affect in voicemail correspondence:
While valence and activation represent the purer emotions - urgency and formality are more complex and more difficult to define in terms of their acoustic correlates.
|
