Learning Tone

University Of Chicago, Chicago IL

Investigators

Abstract

Human languages make crucial use of pitch to convey information. This information ranges from word meaning in languages with lexical tone to syntactic and pragmatic information across a wide range of languages. Furthermore, child language research has suggested that hyper-articulated pitch, loudness, and duration in child-directed speech play a role in bootstrapping intonational as well as lexical and syntactic acquisition. However, despite the fundamental importance of this tone information, computational approaches to speech recognition and processing have largely viewed such pitch variation as a source of noise to be normalized away. This project builds on recent phonetic research that identifies the key role of context in tone realization and pitch variation, explained through mechanisms of maximum rate of pitch change and tonal coarticulation. This research develops a broader-context, articulatorily-motivated model of tone, utilizing a common framework across a range of language and tone typologies including Bantu languages, Chinese dialects, and English. The hyper-articulation of child-directed speech is exploited to identify linguistically relevant variation and to understand its role in tone acquisition. Through unsupervised learning, this work automatically identifies tone and pitch accent in natural speech, while highly leveraging sparse, manually annotated resources for gold-standard evaluation. The improved techniques for modeling and recognition of tone developed in this project will allow computational spoken language understanding systems to more fully exploit the information carried by pitch. These components will also enhance support for language learning through integration of feedback on tone and pitch accent use in a computer-assisted learning system.