中文句法輔助朗讀評分於伴讀型寵物機器人之研究

Abstract

本文的目的為發展伴讀型寵物機器人，輔助教師對於兒童之流暢度評估。 評估的結果會用來做為小孩與伴讀型機器人的回授抑或是教師和家長對於小孩學習情況的掌握因素。首先，將在此伴讀型寵物機器人架設自動語音辨識系統，此語音辨識系統包含了：聲學模型、語言模型和中文句法的階層式隱藏馬可夫模型。聲學模型乃根據人的發聲方式而訓練出特定的模式；語言模型是根據教材中的文章，去計算出字與字之間的關係，最後產生出中文字串；而中文句法的階層式隱藏馬可夫將語音的辨識系統結合中文句法結構樹，使其對於每句辨識結果都會產生到語法樹之葉的節點中。再由各個節點去判斷閱讀時的正確與否。除了朗讀時的準確率外，流暢閱讀的評分條件還包括了：字的間隔時間、閱讀速度、音高、重音與發音。根據這些特徵圖樣，將計算學習者與示範者之間的特徵圖樣之距離。總和以上六種特徵，找母語非中文的人和母語是中文的人來做實驗，驗證此評分系統的可行信，將學習者與示範者的資料做一比對，並探討其結果。最後，將此閱讀流暢度中有問題的指標，轉換成有效的回授給閱讀者以提升其閱讀的成就。

The study investigates a fluency scoring technique for a reading assistance robot. The scoring technique is utilized for the evaluation of oral reading fluency to assist teachers by quantifying children’s reading achievement from children’ reading voices. The scoring of oral reading fluency could be used as a feedback when children are learning and it also can be regarded as a kind of evaluation tool to let the teachers or parents know the learning status of children. An automatic speech recognition system based on acoustic recognizer, language model and Chinese grammar based hierarchical hidden Markov model (CGBHHMM) is established. Acoustic model is trained by human pronunciation. Language model is trained to find the relationship between word and word from elementary school text book materials. CGBHHMM is a statistical model trained by the Chinese grammar tree structure. In the CGBHHMM, each sentence of acoustic syllabus is clustered into phrase production state, and CGBHHMM is then combined with ASR to detect a learner’s word accuracy. Five indicators, read speed, pause duration, pitch, stress and pronunciation, are considered as the features of oral reading fluency (ORF). The distance of ORF indicators is calculated of learners with respect to fluent teachers. These distances of ORF features were compared between fluent readers and foreigners who have learned Chinese for two years. It is verified that the proposed scoring method is effective to detect the fluency differences of fluent and influent readers. For future applications, oral reading fluency is could be used in real time by the assistance robot as feedback instructions to guide children for improving their reading achievement.