Effect of speaker variability on tone perception for L2 learners: Perspectives of machine learning

Mandarin tones 2 and 3 share similar acoustic characteristics, posing a challenge for L2 learners to distinguish. This difficulty is further compounded by both inter- and intra-speaker variability. This study employed machine learning methods to examine the effectiveness of eight acoustic cues for tone 2 and tone 3 classification in speech that were produced by 20 speakers under normal and loud speaking modes (experiment 1). Additionally, we compared the perception between native listeners and medium-to-advanced Thai L2 learners of Mandarin, including perceptual accuracy, cue-weighting strategies, and perceptual space (experiment 2). Results of experiment 1 show that temporal cues are more effective than height-related cues for tone 2 and tone 3 classifications in speech with inter- and intra- speaker variability. In experiment 2, medium-to-advanced Thai L2 learners produced more confusions than native listeners. The ranking of perceptual acoustic cues is generally similar to native listeners, suggesting that these L2 learners have generally mastered the correct cue-weighting strategy for distinguishing between tones 2 and 3. Additional analyses showed that the confusion stems from learners' perceptual biases. Specifically, learners allocate narrower space for tone 3 category than native listeners, with atypical tone 3 exemplars being misidentified as tone 2.