本論文主要是設計一套可用於門禁之語音辨識系統，利用語者辨識技術，判斷輸入聲音是否為核可的使用者之聲音，並結合關鍵詞萃取技術，使系統可辨識出使用者及姓名，且再配合語音合成技術，讓系統不單是純文字的回應，而是模擬人聲之回應，之後經過程式語言包裝，建立一個人機介面的系統，方便使用者操作使用。 因為是門禁系統，需要達到即時或是線上的要求，因此使用到的方法所花費之時間必須考慮，無法將許多方法通通加入，沒辦法讓使用者等待太久才得知結果，所以在方法必須有所篩選，這當然對辨識率有一定程度的影響，但也只能以時間為先決條件，去選擇合適的演算法。在語者辨識部份，經過自行錄製的實驗測試，直接使用使用者的聲音各自建立專屬模型，效果會比經貝氏調適法調適後的模型好。而在關鍵詞部份，因為系統有可新增使用者之功能，所以不可能事先知道使用者姓名，然後針對使用者姓名做模型訓練，改成使用次音節模型，再串成對應的模型，省去各別訓練的時間提高實用性。 從自行測試的實驗結果得知，系統核可使用者人數 38 人，全部測試人數 40 人，有兩個人是模擬仿冒者情況進行測試，語者辨識率 94.9% ，錯誤接受率 0.8% ，關鍵詞辨識率 90.6% ，而平均辨識一句都各自約為 0.5 秒，辨識已可達即時之要求。 The purpose of this thesis is to design a speech access system with speaker recognition technology which can determine whether the input sound of the user voice is valid or not. Combined with keywords spotting technology, the system can identify the name of users. And coupled with text-to-speech technology, the system uses not only a text but also human voice response. System built by Microsoft Foundation Classes (MFC) windows based interface is facilitated for the user to operate. Because access control system needs to meet the requirements of real-time or online, as the result, the consumed time of used methods must take into account because users would not spend much time waiting for results. Therefore, methods must be selective since they affect the recognition rate and time seems to be regarded as the prerequisite element while selecting the appropriate algorithm. There are 40 participants join this test, and there are 38 target users among them, while the other two are imposers. Speaker recognition rate is 94.9%, the false acceptance rate is 0.8%, and the keyword recognition rate is 90.6%. The average recognition sentences are about 0.5 seconds each. Identification has been up to the real-time requirements.
