使用MAT2000語料庫之中文語音辨認

Abstract

在本論文中，我們使用ＭＡＴ2000的語音資料庫來建立國語連續語音之基本辨認系統，首先我們建立了100個右相關聲母與40個韻母模型的基本系統來進行辨認，基本系統辨認率56.65%。接著針對基本系統加以改進，首先，在系統中加入加瑪狀態長度模型，並建立右相關韻母模型來克服聲音耦合的問題，以及使用高斯混合模型替代分割高斯模型來計算狀態觀測機率，使得辨認率提升至67.07%。並嘗試在前處理中加入遞迴式類神經網路與有限狀態機，使得運算速度增進許多，另外也提出一項新的方法來補償語者與通道效應。

In this thesis, a Mandarin speech recognition system using MAT2000 database is constructed. First, 100 final-dependent initial and 40 final HMM models are used to construct the baseline system. An accuracy rate of 56.65% was achieved. Then the Gamma duration model is employed to model the state duration and the Gaussian mixture model is used to model the observation probability instead of Partition Gaussian model. In order to model the co-articulation effect, the right-context-dependent final models are used. The accuracy rate was raised from 56.65% to 67.07% by using the above methods. Besides, the RNN and FSM are proposed to pre-segment the input speech signal. Thus the computation load of the one-stage recognition process can be significant reduced. Finally, a new method is proposed to compensate the speaker/channel effect.