영어 모음 포먼트 군집화 분석

An Analysis of Vowel Clusters Using Formants in Bark.

This study aims to investigate how vowels are categorized into a given number of clusters using an unsupervised machine learning technique called the k-means clustering algorithm. The results of matching clusters to vowel types can explain to what extent the categorizations of vowels are based purely on parameters obtained from the data. The parameters used for the clustering are the first and the second formants in bark. The dataset used in this study is from Hillenbrand et al. (1995). The target vowels were restricted to those produced by adult males. The k was increased stepwise from 3 to 9 to see what phonological features appeared to be active in separating a new cluster from the originating cluster. It was found that the feature [±back] is one of the primary features that differentiates vowel classes and the feature [±high] is used in subsequent separations of the vowel classes. It was also found that high confusability in the perception of vowels is attributable to overlapping distributions of the samples. Vowel duration is critical in clarifying /ɔ/ and /ʌ/, and /ɛ/ and /æ/ as was reported in Hillendbrand et al. (2000). In our study, /æ/ was the second last and /ɔ/ was the last cluster separated from /ɛ, æ/ and /ʌ, ɔ/ clusters, respectively.