Abstract
This paper proposes a method of comprehensively and structurally characterizing the segmental aspect of speech. After building phoneme HMMs from speech samples, distances between any two states of any phoneme HMMs were calculated to make a state-based distance matrix. With Ward's method of hierarchical clustering, a tree diagram of the distance matrix was generated to visualize comprehensive structure embedded in the pronunciation. Using American English (AE) and Japanese English (JE) speech samples, two kinds of trees were drawn. Comparison between the two trees clearly showed the well-known Japanese habits in speaking English. Using the distance matrix of JE, the compatibility between phonetic structure in the pronunciation of JE and lexical structure of the entire vocabulary of AE was estimated based upon Cohort model of isolated word perception.
