電話語音查號系統之改進

Abstract

本論文製作了一個具有簡易對答功能的改良型線上國語語音查號 系統,它使用隱藏式馬可夫模型辨認技術,可辨認13689個電話用戶名稱.研 究主題包含兩個部份,第一部份針對前處理進行語音預切割,以一個遞迴類 神經網路進行語音預切割,將語音分割成聲母,韻母,靜音以及過渡狀態.另 外,在獨立詞辨認過程中加入光束搜尋法,以增進辨認速度.第二部份為利 用改良型的SIFT基頻求取方法,獲得更為可靠的語音基頻軌跡,進而提昇聲 調辨認的結果,並利用聲調訊息,調整獨立詞辨認結果.系統性能採用960語 料測試,在不考慮聲調的情況下,獲得92.4%的獨立詞辨認率,加入聲調辨認 後,辨認率提升為93.89%,辨認速度為一秒鐘語音花費1.25秒的辨認時間. In this thesis, an on-online telephone number inquiry system is implemented on a PC with Dialogic D41/D telephone interface card and a 16-bitSoubd Blaster card. It is an isolated-word recognition system operating underWindows 95. The vocabulary contains 13689 names of companies listed on theyellow page of a local center office. Two main topics are intensively studied.One is the use of a pre-classification scheme and a beam search algorithm toimprovement the recognition speed. The pre- classification scheme pre-classifieseach input frame into one of four broad classes including initial, final, silence,transition, and then sets path constraints to the recognition search. The other topic is to improve the SIFT pitch detection in order to obtain better pitch contourfor tone recognition. Performance of the system was examined by simulations using a database containging 960 utterance uttered by 12 speakers. Recognition rates of 92.4% and 93.89% were obtained, respectively, for the recognizers without and withtone recognition. The Recognition speed was 1.25 second per 1-second speech.