|
Experimental tests of features and partial specification| old_uid | 9468 |
|---|
| title | Experimental tests of features and partial specification |
|---|
| start_date | 2010/12/17 |
|---|
| schedule | 10h30-11h30 |
|---|
| online | no |
|---|
| summary | Representing speech via distinctive features is attractive from a theoretical point of view: such representations let you describe many languages in common terms. But, abstract representations are only valuable if they are also accurate, so we set out to test how well features describe actual tongue positions.
Moreover, there are many feature-based representations in the
literature (which unfortunately means that we are often not using
common terms to describe language), so we set out to find which
ones work best.
The experiment involved obtaining videos of speech with Magnetic
Resonance Imaging, then we built mathematical models of tongue
shapes in terms of phonological features. We found that fully specified feature representations are worse than equivalently complex phoneme models or variants of Harshman and Ladefoged and Goldstein's 1977 representation that represents speech in terms two gradient factors.
On the other hand, the best models we tested turned out to be
partially-specified feature representations. But, among them,
the best model involved a "trinary" filling of unspecified features,
dependent on the phonemes both to the left and to the right. |
|---|
| responsibles | Bel, Welby |
|---|
| |
|