Listen to the beat: A cross-linguistic perspective on the use of stress in segmentation

Abstract

The purpose of this thesis is to investigate the relation of word stress to word segmentation in a cross-linguistic perspective. While many studies have addressed this issue before, the current one takes a typologically broad cross-linguistic approach to the use of edge-aligned word stress in processing. The investigation is concerned with language-specificity, the direction of processing and the language-specific abstract nature of stress as a leading beat. The thesis concludes with an excursion into first language acquisition, regarding the issue of whether word stress can be inferred from the distribution of stress patterns in continuous speech. Word segmentation is the division of continuous speech into words. It is a non-trivial task, since spoken language is fast and words are not divided by pauses, despite listeners’ strong intuition to the contrary. This intuition originates in the listener’s use of language-specific means for segmentation. Word stress is likely one of these. The idea that stress can function to mark word boundaries dates back to a distinction made by Trubetzkoy (1939/1969) and the role of word stress in processing has been tested since some decades later (Taft 1984, Cutler & Norris 1988, and many more). However, the results do not disentangle language-specific from universal segmentation strategies. In the current thesis, a non-word-spotting experiment with a typologically broad selection of languages confirms the Language-Specific Metronome Hypothesis: listeners have unidirectional language-specific stress-based expectations in segmentation. An offline experiment in turn shows that the elimination of time pressure leads listeners to make bidirectional and optimized segmentation decisions, as articulated in the Language-Specific Metrical Grouping hypothesis. Word segmentation without any knowledge of words, as in first language acquisition, has a paradoxical relation to word stress. Word stress could be useful in the task, but how is the word stress system acquired without words? This part of the dissertation is concerned with the question whether it is in principle possible to acquire word stress from unsegmented speech. It is found that statistical relations between adjacent elements fail to capture word stress regularities and the relation of stress to the phrase boundary is more informative, although this differs between languages. Conservatively, it is difficult to acquire the word stress system from unsegmented speech.