Publication Date
4-2018
Abstract
We are interested in enabling a robot to communicate with more natural timings: to take turns more appropriately. LSTM models have sometime been effective for this, but we found that this to be not helpful for some tasks. This technical report we look for factors that may explain this difference, by examining statistically the prosodic feature values in the vicinity of turn shift in the data. We observe that the apparent informativeness of prosodic features varies greatly from one dataset to another.
Comments
Technical Report: UTEP-CS-18-36