Dialect-based speaker classification of Chinese using acoustic features invariant with extra-linguistic factors

Chinese dialects-based speaker classification using modern speech technologies is really a challenge, not only because the situation of Chinese dialects is very complicated, but also because acoustic features of utterances convey dialectal information together with extra-linguistic information such as age, gender, speaker, etc. In this paper, we propose a new speaker classification technique using structural representation of pronunciation, which was originally proposed to remove extra-linguistic information from speech. After collecting dialectal utterances of a selected set of Chinese characters, each speaker is modeled as his/her pronunciation structure. Then, all the dialectal structures are classified based on bottom-up clustering. We also test the proposed method especially in terms of robustness to speaker variability. Here, the utterances of simulated very tall and short speakers are classified to compare the proposed method and the conventional one. All the experimental results show linguistically-reasonable classifications and the high independence of age and gender.