強健式國語連續音節辨認

Abstract

在本論文中, 我們進行以聲母-韻母模型為基礎的電話國語連 續411音節辨認. 研究的主題, 在於建立整合雜訊,通道變化,語者模型之 強健式訓練法與補償式辨認,將PMC以及通道偏移量補償導入訓練的過程 中, 並以Wiener filtering以及通道偏移量之移除得到進乎乾淨語音之模 型參數. 系統性能驗證, 首先採用模擬電話通道之語音測試, 獲得良好的 結果. 最後經電話語音測試, 使用195人的語料訓練及32人的語料測試, 以PMC及通道偏移量補償辨認, 獲得411音節的辨認率為53.96%. In this thesis, a robust training algorithm, which considers the jointcompensation for noise and channel/speaker effects, is proposed for Mandarintelephone speech recognition. It incorporates the PMC method and a bias compensationmethod into the segmentation stage of the training process in order to adapt thecurrent HMM models to the noisy environment. It also employs the Wiener filteringand the SBR method in the model updating stage of the training process in order toobtain clean speech for updating the current HMM models. Effectiveness of the proposed algorithm was first confirmed by simulations using a database containing many utterancesof simulated telephone speech. It was then tested using a real telephone-speechdatabase generated by 227 speakers including 195 for training and 32 for testing. By using the PMC method and the SBR method in the recognition test, a base-syllablerecognition rate of 53.96% was achieved.