Confirmatory Questions Model

This is a more contstrained, more predictive, stronger model than bare Stem-ML, and captures the regularities of the speech very compactly.

The model consists of a Stem-ML stress tag on each word. All words get a stress tag derived from the same template, except the single word that is being confirmed (here called the accented word). All instances of the accented word share the same template. In other words, we assume there are two classes of accent in the data: one class for the number needing confirmation, and one class for everything else.

There is also one stress tag placed at the beginning, to represent an initial boundary tone, and one placed at the end to represent the final boundary tone. All the initial boundary tones share one template, and all the final boundary tones share another template.

A template is defined by 5 (3 for boundary tones) pitch values, spaced across its scope. It is merely stretched (in time) and scaled (changing its pitch range) to describe all words which use that template. Each template has a Stem-ML type parameter, which controls how it interacts with its neighbors. Templates also have an atype parameter, which controls how the template scaling depends on each word's strength.



Subsections
Greg Kochanski, Chilin Shih 2002-08-03