Predicting confidence and doubt in accented speakers: Human perception and machine learning experiments

Xiaoming Jiang, Marc Pell


Speech prosody provides salient and reliable cues to facilitate social communication. What computational mechanism underlies social judgment towards “out-group” speakers is unclear. This paper focused on Speaker Confidence, a factor affecting one’s trustworthiness, persuasiveness and feeling of (un)knowing, and Speaker Accent, a factor marking one’s identity. We demonstrate that native Canadian-English listeners can recognize confident and doubtful expressions in foreign- and regional-accented speakers. A stronger impression of confidence was shown towards the native speakers. The acoustic analysis demonstrated that speakers systematically varied the mean fundamental frequency to indicate confident and doubt regardless of accent. The out-group speakers varied more on intensity height and variation to achieve certain level of confidence. Machine learning experiments showed above-chance accuracies in all accents to classify vocal expression based on global acoustic cues, highlighting the role of acoustic regularities at utterance level in confidence encoding. Moreover, the classification rate was higher when the model trained in native accent was tested on the native than the regional accent, highlighting an in-group bias of predicting novel vocal expression of confidence from acoustic cues. These findings lend support to the dialect theory of vocal expression recognition while demonstrating a computational mechanism underlying inter-cultural/inter-group confidence perception via speech prosody.


 DOI: 10.21437/SpeechProsody.2018-55

Cite as: Jiang, X., Pell, M. (2018) Predicting confidence and doubt in accented speakers: Human perception and machine learning experiments. Proc. 9th International Conference on Speech Prosody 2018, 269-273, DOI: 10.21437/SpeechProsody.2018-55.


@inproceedings{Jiang2018,
  author={Xiaoming Jiang and Marc Pell},
  title={Predicting confidence and doubt in accented speakers: Human perception and machine learning experiments},
  year=2018,
  booktitle={Proc. 9th International Conference on Speech Prosody 2018},
  pages={269--273},
  doi={10.21437/SpeechProsody.2018-55},
  url={http://dx.doi.org/10.21437/SpeechProsody.2018-55}
}