Towards Immediate Backchannel Generation Using Attention-Based Early Prediction Model

Backchannel feedback from the spoken dialogue system makes the human-machine interaction more sophisticated. To predict suitable timing and forms, backchannel prediction technology has been studied. Most studies have combined acoustic and lexical features into the model for better prediction. However, extracting lexical features leads to a delay caused by the automatic speech recognition (ASR) process. To make accurate predictions on the basis of delayed ASR outputs, we propose early prediction for backchannel opportunity and backchannel category based on attention-based LSTM mechanisms. The loss is calculated with a weighting value that gradually increases when a sequence is closer to a suitable response timing. The proposed backchannel prediction uses a two-step approach that first detects a backchannel opportunity and then predicts a backchannel category. Evaluation results show that the early prediction model can predict a backchannel opportunity and category better than the current state-of-the-art algorithm even under a 2.0-second ASR delay condition.