Semi-automatic labeling of intonation in ToBI framework

Diana Stojanovic, Linguistics

Yohei Sakata, Linguistics

Intonation is an important part of speech into which semantic and pragmatic information is encoded. Different intonation patterns will distinguish a question from a statement and a hesitant from a confident response. The process of marking intonation patterns is usually performed manually, by trained researchers, based on the auditory impressions. In studies on intonation dealing with large quantities of data, an automatic prosodic labeling system is highly desired. Several attempt have been made to develop such a system (Mertens 2004, Hirst 1999, Campbell 1996, Taylor 1995).

    The goal of this paper is to describe a semi-automatic program that marks the pitch accent of American English in ToBI (Tones and Break Indices) framework, based on the acoustic measurements of pitch. A set of audio samples recorded by a single female speaker (Beckman, 1994) was used to train and test the program. Acoustically similar types of accents were grouped into equivalent labels to achieve accuracy level of 61- 78%.

    The applications of such a program are multifold. First, it can be used by researchers for semi-automatic labeling of intonation, thus saving substantial amount of researcher’s time in larges studies. Second, it can be used as an educational tool, as the interactive version of the program helps students master the intonation labeling technique.

    The presentation will involve demonstration of the program, as well as the discussion of challenges encountered during development process and the directions in which this project may evolve in the future.