Structural analysis of Chinese dialects and its experimental application to pronunciation assessment

摘要

The current situation of Chinese dialects is very complicated, not only because there are hundred kinds of dialects, sub-dialects and sub-sub-dialects, but also because many dialect speakers are speaking Mandarin with different regional accents affected by their dialects. If dialect utterances are represented acoustically, their acoustic features depend both on dialectal linguistic features and extra-linguistic features such as the age, gender, micro phone, etc. Then, the automatic pronunciation assessment of Chinese dialects becomes a challenging task. In this paper, we propose to analyze the dialect pronunciation by the structural representation, which was originally used to remove extra-linguistic information from speech. After collecting the utterances of some selected Chinese written characters, the pronunciation structures are built for dialect speakers and their pronunciations are compared by the distances among the structures. Then with the dialect utterances of 16 native dialect speakers, some experi ments are carried out to analyze their pronunciations using the structures and linguistically-reasonable results are obtained. At last, by comparing the pronunciation structures of dialect speakers with those of two standard Man darin speakers, the pronunciation similarity orders of the characters are obtained. Through these experiments, our approach is proved again that it works well in extracting the linguistic features from dialect utterances and shows high independence of the extra-linguistic features.%中国語方言は、地理的社会的言語的要因により非常に複雑な状況にある。中国には数百種類もの方言、下位方言、更なる下位方言があり、方言話者が標準語を話す場合にはそれらの影響を強く受ける。通常、音声の音響的特徴はスペクトルによって表されるが、これには方言的特徴の他に、性差、年齢差、個人差などの非言語的特徴も含まれているため、中国語方言話者に対するの自動発音評価は容易ではない。本研究では、音声の構造的表象を用いて非言語的情報を効果的に除去した後になまりのある発音を評価する手法に関する実験的検討を行なう。この手法では、まず漢字音セットに対する方言の録音を行った後、話者ごとに発音構造を構築する。そして他の話者との構造間距離を比較することで、発音を評価する。本稿では同じ漢字の方言発音と標準語の発音を比べることで、発音の類似度を算出し、それが性別に対して高い頑健性をもつことを示した。