Exploiting loudness dynamics in stochastic models of turn-taking

摘要

Stochastic turn-taking models have traditionally been implemented as N-grams, which condition predictions on recent binary-valued speech/non-speech contours. The current work re-implements this function using feed-forward neural networks, capable of accepting binary- as well as continuous-valued features; performance is shown to asymptotically approach that of the N-gram baseline as model complexity increases. The conditioning context is then extended to leverage loudness contours. Experiments indicate that the additional sensitivity to loudness considerably decreases average cross entropy rates on unseen data, by 0.03 bits per framing interval of 100 ms. This reduction is shown to make loudness-sensitive conversants capable of better predictions, with attention memory requirements at least 5 times smaller and responsiveness latency at least 10 times shorter than the loudness-insensitive baseline.