Measuring Prosodic Transfer in Vector Space by Weighted Tonal Events

Prosodic transfer plays a significant role in forming stereo-typed foreign accent, so measuring the transfer of non-native speakers is of great importance to language learning. Recent researches are trying to build intonation models to describe the transfer using f0 and break differences between the target and the gold standard. This paper advances a method to compute prosodic distance in vector space by the weights of each prosodic event abstraction using ToBI labels produced by AuToBI. This is an attempt to evaluate prosodic transfer in a highly generalized level that could match cognitive reality. N-gram models of weighted tonal events exhibits separability of Ningbo speakers from US, and among learners of varying prosodic capacity in high dimensional vector space. Thus, it proves a right direction to measure prosodic transfer with discrete representations of intonation.