Fluency Profiling System: an automated system for analyzing the temporal properties of speech

摘要

The temporal characteristics of speech can be captured by examining the distributions of the durations of measurable speech components, namely speech segment durations and pause durations. However, several barriers prevent the easy analysis of pause durations: The first problem is that natural speech is noisy, and although recording contrived speech minimizes this problem, it also discards diagnostic information about cognitive processes inherent in the longer pauses associated with natural speech. The second issue concerns setting the distribution threshold, and consists of the problem of appropriately classifying pause segments as either short pauses reflecting articulation or long pauses reflecting cognitive processing, while minimizing the overall classification error rate. This article describes a fully automated system for determining the locations of speech–pause transitions and estimating the temporal parameters of both speech and pause distributions in natural speech. We use the properties of Gaussian mixture models at several stages of the analysis, in order to identify theoretical components of the data distributions, to classify speech components, to compute durations, and to calculate the relevant statistics.