Modeling phrasing and prominence using deep recurrent learning

Abstract

Models for the prediction of prosodic events, such as pitch accents and phrasal boundaries, often rely on machine learning models that combine a set of input features aggregated over a finite, and usually short, number of observations to model context. Dynamic models go a step further by explicitly incorporating a model of state sequence, but even then, many practical implementations are limited to a low-order finite-state machine. This Markovian assumption, however, does not properly address the interaction between short- and long-term contextual factors that is known to affect the realization and placement of these prosodic events. Bidirectional Recurrent Neural Networks (BiRNNs) are a class of models that overcome this limitation by predicting the outputs as a function of a state variable that accumulates information over the entire input sequence, and by stacking several layers to form a deep architecture able to extract more structure from the input features. These models have already demonstrated state-of-the-art performance on some prosodic regression tasks. In this work we examine a new application of BiRNNs to the task of classifying categorical prosodic events, and demonstrate that they outperform baseline systems.