噪音環境下的漢語語音辨識

Abstract

語音辨識系統的辨識率受到環境噪音很大的影響。在克服噪音影響的方法 中，語音特徵修正和語音模型修正是兩大類方式。語音特徵修正方式嘗試 將噪音語音修正成乾淨語音，而語音模型修正則嘗試將乾淨語音模型修正 為噪音語音模型。本論文比較這兩大類方式中的幾種方法。在語音特徵修 正方式中我們討論頻譜消去法和線性特徵轉換法。在語音模型修正方式中 我們研究最小平方差模型平均值修正法，最大機率線性回歸法，最小分類 錯誤線性回歸法，投射法和對等模型合併法。在本論文我們實驗這幾種方 法在八種不同噪音環境下的辨識率，並研究彼此的差異性。 The recognition rate of a recognition system degrades seriously in a noisy environment. In the literature, there have been various methods directed to improve the recognition rate. Among them, two approaches deserves special noting: speech feature compensation and model compensation. Speech feature compensation tries to provide compensation method for the speech features while model compensation provides the suitable modification on the model parameters. The main purpose of this thesis is to compare several speech feature compensation methods and speech model compensation methods. Among the feature compensation methods, we consider the spectral subtraction method and the linear feature mapping methods. Among the model compensation methods, we consider the model mean adaptation with least square error, the maximum likelihood linear regression, the minimum classification error linear regression, projection and the parallel model combination. This thesis compares the performance of the seven methods through eight types of noise environments and shows the dependence of each method with various parameters.