Prominence Detection without Syllabic Segmentation

Philippe Martin, Paris Diderot

Detection of prominence, whether automatically or manually through perception tests, is pivotal in the interpretation of data in a prosodic theoretical framework. This is particularly true for French, where phonologically stressable syllables are not necessarily stressed. To assert a prominence character to syllables is mandatory to evaluate prosodic theories, especially those which predict the phonetic features of melodic contours (rise, fall, height, etc.) located on those syllables. Some algorithms are already available to detect prominent syllables automatically, but most involve a precise segmentation of speech into syllables, vowels and consonants, a task which generally requires a reasonable good quality of recording, exempt from background noise and echo. In order to avoid the problematic segmentation into phonetic units, we propose here an algorithm for prominence detection operating differently and based on readily available phonetic properties of speech, at the exemption of spectral properties.