Effects of variance and input distribution on the training of L2 learners' tone categorization

Abstract

Recent psycholinguistic findings showed that (a) a multi-modal phonetic training paradigm that encodes visual, interactive information is more effective in training L2 learners' perception of novel categories, (b) decreasing the acoustic variance of a phonetic dimension allows the learners to more effectively shift the perceptual weight towards this dimension, and (c) using an implicit word learning task in which the words are contrasted with different lexical tones improves naïve listeners' categorization of Mandarin Chinese tones. This dissertation investigates the effectiveness of video game training, variance manipulation and high variability training in the context of implicit word learning, in which American English speakers without any tone language experience learn four Mandarin Chinese tones by playing a video game. A video game was created in which each of four different animals is associated with a Chinese tone. The task for the participants is to select each animal's favorite food to feed it. At the beginning of the game, each animal is clearly visible. As the game progresses, the images of the animals become more and more vague and eventually visually indistinguishable. However, the four Chinese tones associated with the animals are played all through the game. Thus, the participants need to depend on the auditory information in order to clear the difficult levels. In terms of the training stimuli, the tone tokens were manipulated to have a greater variance on the pitch height dimension, but a smaller variance on the pitch direction dimension, in order to shift the English listeners' perception to pitch direction, a dimension that native Chinese listeners crucially rely on. A variety of pretests and posttests were used to investigate both the English speakers' perception of the tones and their weighting of the acoustic dimensions. These training stimuli were compared to other types of training stimuli used in the literature, such as the high variability natural stimuli and tones embedded in non-minimal pairs. A group of native English speakers was used as the control group without any tone input. A native control group was also included. The video game training for each speaker consisted of four 30-minute sessions on four different days, and 60 participants (including both the non-native control and native control group) participated in the experiments. The crucial findings in the study include (1) all naïve listeners in the training condition successfully associated lexical tones with different animals without any explicit feedback after only 2 hours of training; (2) both the resynthesized stimuli with smaller variance on pitch direction and the multi-talker stimuli allowed native English speakers to shift their cue-weighting toward pitch direction and the multi-talker stimuli were more robust in terms of shifting the cue-weighting despite their more heterogeneous distribution in the acoustic space; (3) the multi-talker training allowed for better generalization as the trainees in multi-talker training identified the tones produced by new talkers better than trainees in other conditions; (4) there was a main effect of tone on tone identification and the falling tone was the most challenging one; (5) there is a correlation between cue-weighting and the tone discrimination performance before and after the training; (6) due to individual variability, individuals differed in terms of the amount of tone input they received during the video game training and the number of tone tokens was a significant predictor for the sensitivity to tones calculated as d'. Overall, the study showed an effect of talker variability and variances of multidimensional acoustic space on English speakers' cue-weighting for tone perception and their tone categorization.