最近幾年開始,世界各地開始有一股學習中文的熱潮。各種中文學習的書籍或是課程不斷地問市,使得中文學習的市場非常的熱絡。相較於有關英文學習的電腦軟體隨處可見,有關中文學習的電腦軟體在市面上卻不多見,可隨身攜帶學習的中文口語學習機更是非常少見。本實驗室已將基於定點數的語音辨識系統,修改成為能夠在嵌入式系統上運作,且基於定點數運算的語音評分系統。但此語音評分系統的得分結果並不是十分客觀,且有評分執行速度太慢等問題。我們希望能針對此系統的問題進行研究,並且找出可行的辦法,希望能夠使得執行速度加快,且評分的結果能夠更加正確。 本論文以基於在嵌入式系統平台上運作的定點數運算之語音評分系統,觀察並研究評分系統所得出的評分錯誤結果資料,並根據這些觀察結果,設定條件來進行分數的調整,希望經過調整後的得分,能夠更加客觀而讓使用者能夠接受,使得系統的評分準確度能夠更高。另外在不改變系統核心下,加入額外的處理,使得評分的執行速度能夠加快。
This thesis explores the possibility of improving the performance of our Mandarin speech assessment systems on 32-bit fix-point platform. For improving efficiency, we have proposed several methods for reduce computation, such that the response time of the system can be as short as possible without degrading its performance. For improving effectiveness, we have also proposed four score-correction rules that can be used to give a more consistent scores of speech assessment. We have implemented these methods and rules on a PMP (personal media player) based Mandarin speech assessment system with satisfactory performance both in efficiency and effectiveness.