不特定語者之中國方言自動辨識

Abstract

本論文之研究目的在於發展一種不特定語者、不特定詞彙 之方言自 動辨識系統，特別是針對臺灣境內的三種主要方言— 北京話、河洛話及 客家話來進行辨識。研究主題可依辨識方法 區分為三個部份，第一部份 是從音韻學的觀點來探討方言辨識 的問題。我們將輸入語句轉換表示為 一連串的粗分音類序列， 亦即進行音類分辨，再藉由這些音類連接的變 化資訊作為鑑別 方言的依據。其中音類分辨器之製作是先分割單字音中 的聲母 和韻母部份，再分別由所屬模型進行辨認。而方言辨識主體則 採 用隱藏式馬可夫模型及遞迴式神經網路這兩部份各別進行， 經人工標示 音類序列測試結果，前者之最佳辨識率為81.11%， 而後者優於前者，可 達92.33%。在第二部份中，我們改站在聲 學的角度，直接從輸入語句中 擷取特徵參數來進行鑑定，經實 際語音測試，最佳辨識率為84.44%。至 於第三部份，則是結合 了前述兩部份之鑑別模式，將音類序列的變化資 訊融入聲學辨 識系統設計中，同樣以實際語音進行測試，獲得辨識率大 幅提 升至96.94%，為方言辨識相關研究提供了一套非常可行的方法 。 As a part of multilingual spoken language system, realiabletechniques are needed to identify various dialects in order toroute the user to the appropriate human or information accesssystem. This work is aimed to develop an automatic dialect id-entification that takes speaker-independent, context- independ-ent utterances as input and produces a dialect hypothesis asoutput. The system has been trained to recognize three Chinesedialects (Mandarin, Holo, and Hakka), but could be extended easily to include other dialects as well. We begin by studyinga dialect identifier in conjection with broad phonetic classi-fication of speech. It is expected that dialects differ from each other with respect to their typical sequential statisticsof broad phonetic classes. Such phonotactic cues can be well modeled by either using hidden Markov models or simple recurr-ent neural networks. For purpose of comparison, we have alsoconstructed an identifier based on the stochastic models of short-term acoustics in each dialects. By incorporation acousticand phonotactic information, we reported the benefits of a newhybrid identification scheme that is capable of achieving thehighest accurancy rate of up to 96.94%.