Speech Science Forum - Nadine Lavan & Gwen Brekelmans (QMUL)

Abstract:

High variability phonetic training (HVPT) has become a standard paradigm in teaching novel L2 contrasts since the publication of two seminal studies (Logan et al., 1991; Lively et al., 1993). In HVPT speech contrasts are learned via multiple-talker input, where more varied input should aid generalization by focussing the learner on diagnostic acoustic features, whilst discounting irrelevant cues. However, these studies only include small samples (N=6), with few studies since having directly compared high and low variability input.

We therefore conducted a large-scale replication (N=166) of these studies as a registered report, to establish whether speech perception is better after phonetic training including multiple talkers (high-variability) vs. a single talker (low-variability). Native speakers of Japanese received either high- or low-variability training on the non-native contrast /r/-/l/ by learning to discriminate between English minimal pairs (“lock”-“rock”).

We saw clear learning during training in both high- and low-variability. However, our study found no evidence for a high-variability benefit: Bayes Factor analyses show that the evidence for a high-variability benefit was ambiguous, suggesting a very small effect at best. Our findings raise questions about the smallest effect size of interest for phonetic training, and have important implications for our understanding of speech perception.