Testing low-level speech features against human perceptual data

Abstract: 

Listeners generalize across utterances from different speakers, dialects, listening conditions, and contexts.  I show how a cognitive model that operates over speech corpora can be used to evaluate hypotheses about the dimensions of speech that guide these generalizations.  Simulations show that representations that are normalized across speakers predict human discrimination data better than unnormalized representations, mirroring previous findings.  The model also reveals differences across normalization methods in how well each predicts human data.  These results indicate that cognitive modeling can be used to quantitatively evaluate different representations of speech, yielding consistent and interpretable results.

Biography:

Please visit:
http://ling.umd.edu/~nhf/