Distributional Learning of Lexical Tones: A Comparison of Attended vs. Unattended Listening

This study examines whether non-tone language listeners can acquire lexical tone categories distributionally and whether attention in the training phase modulates the effect of distributional learning. Native Australian English listeners were trained on a Thai lexical tone minimal pair and their performance was assessed using a discrimination task before and after training. During Training, participants either heard a Unimodal distribution that would induce a single central category, which should hinder their discrimination of that minimal pair, or a Bimodal distribution that would induce two separate categories that should facilitate their discrimination. The participants either heard the distribution passively (Experiments 1A and 1B) or performed a cover task during training designed to encourage auditory attention to the entire distribution (Experiment 2). In passive listening (Experiments 1A and 1B), results indicated no effect of distributional learning: the Bimodal group did not outperform the Unimodal group in discriminating the Thai tone minimal pairs. Moreover, both Unimodal and Bimodal groups improved above chance on most test aspects from Pretest to Posttest. However, when participants' auditory attention was encouraged using the cover task (Experiment 2), distributional learning was found: the Bimodal group outperformed the Unimodal group on a novel test syllable minimal pair at Posttest relative to at Pretest. Furthermore, the Bimodal group showed above-chance improvement from Pretest to Posttest on three test aspects, while the Unimodal group only showed above-chance improvement on one test aspect. These results suggest that non-tone language listeners are able to learn lexical tones distributionally but only when auditory attention is encouraged in the acquisition phase. This implies that distributional learning of lexical tones is more readily induced when participants attend carefully during training, presumably because they are better able to compute the relevant statistics of the distribution.