Assessing Acoustic and Articulatory Dimensions of Speech Motor Adaptation with Random Forests

Eugen Klein, Jana Brunner, Phil Hoole


Although most modern theories of speech production assume that representations of speech sounds are multidimensional encompassing acoustic and articulatory information, speech motor learning studies which assess the degree of adaptation in both dimensions are few and far between. In the current paper, we present an auditory perturbation study of German sibilant [s] in which speakers’ audio and articulatory movements were recorded by means of electromagnetic articulography. Random Forest, a supervised learning algorithm, was employed to classify speakers’ responses produced under unaltered or perturbed feedback based either on acoustic or articulatory parameters. Preliminary results demonstrate that while classification accuracy increases in the acoustic dimension as the perturbation session goes on, the classification accuracy in the articulatory dimension, although overall higher, remains approximately at the same level. This suggests that the adaptation process is characterized by active exploration of the articulatory space which is guided by speakers’ auditory feedback.


 DOI: 10.21437/Interspeech.2019-1812

Cite as: Klein, E., Brunner, J., Hoole, P. (2019) Assessing Acoustic and Articulatory Dimensions of Speech Motor Adaptation with Random Forests. Proc. Interspeech 2019, 899-903, DOI: 10.21437/Interspeech.2019-1812.


@inproceedings{Klein2019,
  author={Eugen Klein and Jana Brunner and Phil Hoole},
  title={{Assessing Acoustic and Articulatory Dimensions of Speech Motor Adaptation with Random Forests}},
  year=2019,
  booktitle={Proc. Interspeech 2019},
  pages={899--903},
  doi={10.21437/Interspeech.2019-1812},
  url={http://dx.doi.org/10.21437/Interspeech.2019-1812}
}