Automatic pronunciation error detecting of Chinese using SVM with structural features

摘要

外国語学習者の発声には多くの発音誤りが観測される。彼らの学習を支援するCALL(Computer-Aided Language Learning)システムの構築には,発音誤り検出技術が必須である。本研究では中国語を学ぶ日本人学習者 を対象に，彼らの典型的な音素発音誤りの検出について検討した。音素誤りラベリングが施された学習者音声デー タベースの構築には多大な労力が必要となるため，ここでは母語話者音声データを操作して，人工的に，発音誤り音声データを準備した。まず,日本人にとって発音が難しい8つの音素を定義した。次に，これらの音素に関して書き起こしを操作することで，発音誤りが混入した音声データを作成した。更には故意に誤って発音した音声につ いても収録した。これらのデータを用い，GOP (Goodness of Pronunciation), LR (Likelihood Ratio),及び構造的特徴 を用いたSVM (Support Vector Machine)の三種類の方法を検討した。実験の結果，構造的特徴を用いたSVMが他の 二種類の方法よりも非常に高い精度で誤り検出ができることが示された。%Pronunciation errors are often made by learners of a foreign language. To build a CALL (Computer-Aided Language Learning) system to support them, automatic error detection is an essential technique. In this study, Japanese learners of Chinese are focused on and automatic detection of their typical and frequent phoneme production errors is investigated. Due to difficulty of preparing labels of the phoneme errors found in real learners' data, we prepare native utterances including artificial phoneme errors. First, the target phonemes, which often appear to be problematic for Japanese learners to pronounce correctly, are defined through discussion with teachers. Then, phoneme production errors are created artificially by changing transcripts of native utterances and we further ask native speakers to read sentences with specific phoneme production errors.Three methods of GOP (Goodness of Pronunciation), LR (Likelihood Ratio), and SVM (Support Vector Machine) with structural features are compared under the task of phoneme error detection. Here, GOP-based error detection is done by using the threshold optimized through development data. Results show that SVM with structural features performs better than both of the GOP-based and LR-based baseline methods.