經電話通道之國語連續411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，研究主題包含兩個部份，第一部份係針對語者及通道效應，使用 三種方法對語者及通道正規化，降低語者及通道對辨認之負面影響；第二 部份係探討連續語音的連音效應，使用前後文右相關韻母模型來解決此問 題。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，獲 得411音節的辨認率為54.02%。 In this thesis, an initial-final HMM based method for continuous telephone Mandarin 411-base-syllable recognition is discussed. Two main topics are intensively studied. They are the speaker/channel effect and the coarticulation effect. Three methods are used to compensate the speaker/channel effect on degrading the recognition performance. And a method using 200 right-context-dependent final models is adopted to conquer the coarticulation effect. Performance of the proposed method was examined by simulations using a real telephone-speech database uttered by 227 speakers including 195 for training and 32 for testing. An accuracy rate of 54.02% was achieved.