TL;DR
This paper investigates how different experimental methods (ABX vs. AXB tasks) and reaction time measurements affect our understanding of how Thai listeners perceive the similarity between consonant sounds. They found that AXB tasks generally lead to better accuracy and are preferred by listeners, suggesting it might be a more robust method for assessing perceptual representations, especially when dealing with subtle phonetic distinctions.
What Problem Does This Paper Address?
Understanding how humans perceive and distinguish between sounds is fundamental to speech processing and linguistics. Researchers often use psychophysical tasks to measure this perceptual sensitivity. However, the choice of experimental method can significantly influence the results, potentially leading to different conclusions about how listeners represent and differentiate sounds. This paper focuses on Thai initial consonants, a language with a rich phonemic inventory. Previous work has proposed a perceptual space for these consonants, but the authors wanted to rigorously test how different experimental methodologies, specifically the ABX and AXB discrimination tasks, would impact the assessment of perceptual similarity. They also aimed to explore the utility of reaction time (RT) as a complementary measure to accuracy, hypothesizing that RT could reveal finer-grained perceptual distinctions. The core problem is to determine if these methodological choices introduce biases or reveal different aspects of perceptual representation, and to validate their proposed perceptual space of Thai consonants.
Methods
The study employed two standard psychophysical discrimination tasks: ABX and AXB. In both tasks, listeners were presented with three stimuli (A, B, and X) and asked to determine if X was more similar to A or B. The key difference lies in the order of presentation: ABX presents A and B first, followed by X, while AXB presents A, then X, then B. This difference is hypothesized to affect working memory load.
Thirty pairs of Thai initial consonants were selected, categorized into three levels of perceptual similarity: highly similar, moderately similar, and clearly distinct, based on a previously established perceptual space. These consonant pairs were embedded in rhyming words and presented to listeners in the presence of additive white Gaussian noise at a -12dB SNR.
In addition to accuracy (percent correct responses, p(C)), reaction times (RTs) were meticulously recorded for each correct response. RTs were analyzed after removing outliers (values more than two standard deviations from the mean) to ensure robustness. Statistical analyses, including paired t-tests, were used to compare performance between the ABX and AXB tasks, and across the different similarity groups. Listeners also provided preference ratings for the tasks.
Datasets
The study did not introduce a new dataset. Instead, it utilized a set of 30 rhyming word pairs containing Thai initial consonants, selected from previous research ([1], [10]). These pairs were chosen to represent varying degrees of perceptual similarity. The stimuli were recorded by a native Thai speaker and then corrupted with additive white Gaussian noise.
Key Insights and Ideas
- AXB Task Advantage: The AXB task consistently yielded higher accuracy and was preferred by the majority of listeners, suggesting it might be a more sensitive or less demanding method for assessing perceptual similarity compared to ABX.
- RT as a Complementary Measure: While accuracy is crucial, reaction times provided additional insights. Although not always statistically significant, RT patterns often mirrored accuracy trends, indicating that RT can indeed reflect perceptual sensitivity and decision difficulty.
- Perceptual Space Validation: The results largely supported the predictions derived from the previously proposed perceptual space of Thai initial consonants. Consonants within the same cluster and with small perceptual distances were indeed the hardest to discriminate (lowest p(C), highest RT).
- Methodological Nuances: The study highlights that methodological choices matter. The difference in stimulus presentation order between ABX and AXB can lead to observable differences in performance, even when the underlying perceptual distinctions are the same.
- Potential for Perceptual Boundary Insights: The observed irregularities in RTs for certain "moderately similar" consonant pairs (specifically, those from different clusters but with small perceptual distance) hint at complex perceptual processes and potential sensitivity to phoneme-cluster boundaries.
Key Results
- Accuracy: Listeners performed significantly better in the AXB task (94.6% correct) compared to the ABX task (88.2% correct).
- Preference: 73% of participants preferred the AXB task over the ABX task.
- Similarity Levels: Across both tasks, accuracy was highest for clearly distinct consonant pairs (Group III), intermediate for moderately similar pairs (Group II), and lowest for highly similar pairs (Group I). These differences were statistically significant.
- Reaction Time: While AXB had slightly higher average RTs than ABX, this difference was not statistically significant. RTs generally followed the same pattern as accuracy across similarity levels, with longer RTs for more difficult discriminations.
- Group II Sub-analysis: Consonant pairs from different clusters but with small perceptual distance (Type 1) showed more fluctuation in both accuracy and RT compared to pairs from the same cluster but with larger distances (Type 2).
Limitations and Future Work
- Sample Size: The study involved 22 participants, which is a reasonable number for psychophysical experiments, but larger samples could further strengthen the statistical power.
- Noise Level: The chosen SNR of -12dB was found to be optimal, avoiding ceiling effects at higher SNRs and near-chance performance at lower SNRs. However, exploring performance at other noise levels could provide a more comprehensive picture.
- Future Work: The authors are actively developing an adaptive testing method using real speech to detect hearing deficits. They plan to integrate their Thai rhyming word stimuli and the AXB task into this adaptive system to not only identify hearing problems but also to quantify the degree of hearing difficulty for specific speech sounds. This adaptive approach aims to provide a more personalized and efficient assessment tool.
Summary updated