Automatic Prediction of Confidence Level from Children’s Oral Reading Recordings

Kamini Sabu, Preeti Rao


Perceived speaker confidence or certainty has been found to correlate with lexical and acoustic-prosodic features in the spontaneous speech of children interacting with an adult. We investigate the prediction of confidence in the context of oral reading of stories by children with good word recognition skills where we must rely purely on prosodic features. We report a dataset of oral reading recordings that has been manually rated for confidence at the level of text paragraphs of 50–70 words. Several acoustic features computed at different time scales are evaluated via a trained classifier for the prediction of the subjective ratings. Features based on pausing, pitch and speech rate are found to be important predictors of perceived confidence. Also it is seen that the ratings are influenced by signal properties computed across the utterance. When trained on recordings with strong rater agreement, the system predicts low confidence readers with an F-score of 0.70.


 DOI: 10.21437/Interspeech.2020-2276

Cite as: Sabu, K., Rao, P. (2020) Automatic Prediction of Confidence Level from Children’s Oral Reading Recordings. Proc. Interspeech 2020, 3141-3145, DOI: 10.21437/Interspeech.2020-2276.


@inproceedings{Sabu2020,
  author={Kamini Sabu and Preeti Rao},
  title={{Automatic Prediction of Confidence Level from Children’s Oral Reading Recordings}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={3141--3145},
  doi={10.21437/Interspeech.2020-2276},
  url={http://dx.doi.org/10.21437/Interspeech.2020-2276}
}