Effective Acoustic Cue Learning Is Not Just Statistical, It Is Discriminative

Jessie S. Nixon

A growing statistical learning literature suggests that listeners extract statistical information from the linguistic environment. However, distributional frequency may be insufficient for important but relatively low-frequency cues. Acquisition of linguistic knowledge may rely not merely on co-occurrences but on predictive relationships between cues and their outcomes. The present study investigates effects of predictive temporal cue structure on acquisition of a non-native acoustic cue dimension. During training, native English speakers saw coloured shape objects and heard spoken Min Chinese words with six different lexical tones. Tones were the only reliable cue to identifying the associated object. Words also contained a salient cue that did not discriminate between objects. Three tones occurred with high-frequency and three with low-frequency in training. The critical manipulation was the presentation order: either words, containing complex cue structure, preceded object outcomes (discriminative order) or objects preceded words (non-discriminative order). Generalised linear mixed models showed accuracy was significantly higher in the discriminative order than the non-discriminative order. These results demonstrate that predictive cue structure can facilitate acquisition of a non-native cue dimension. Feedback from prediction error drives learners to ignore salient non-discriminative cues and effectively learn to use the target cue dimension.

 DOI: 10.21437/Interspeech.2018-1024

Cite as: Nixon, J.S. (2018) Effective Acoustic Cue Learning Is Not Just Statistical, It Is Discriminative. Proc. Interspeech 2018, 1447-1451, DOI: 10.21437/Interspeech.2018-1024.

  author={Jessie S. Nixon},
  title={Effective Acoustic Cue Learning Is Not Just Statistical, It Is Discriminative},
  booktitle={Proc. Interspeech 2018},