本論文的目標在於設計一個自動系統,能夠判斷演唱歌詞之正確與否,進而做為歌唱評分的一項依據。我們先利用一個現有的語句確認系統來評估所給定的歌唱訊號是否符合所定義的演唱歌詞,亦即唱詞確認。但結果發現這種以語音資料所建立的語句確認系統並不適合用來處理歌唱訊號,意即系統無法區隔唱對歌詞與唱錯歌詞的訊號。探究其主因在於演唱時常因配合旋律的關係而將母音拉長,造成歌唱訊號與語音訊號存在明顯的差異。為了解決這種的問題,本論文提出母音壓縮法及母音裁剪法來改善系統。主要概念是將被拉長的母音變短,使之接近於語音訊號。經實驗測試,修改母音後可降低唱詞確認的錯誤率約2%-3%,使唱詞確認的效能趨向於說話語句確認的效能。
This thesis aims to develop an automatic system for accessing if the lyrics sung by a performer is correct or not, thereby providing a clue for singing skill evaluation. Our basis strategy is to use a well-established speech utterance verification system to determine if the sung lyrics match the given textual lyrics, which can be considered as a task of sung lyrics verification. However, our experiment results show that a speech utterance verification system cannot handle singing data well, mainly because of the significant differences between singing and speech. One of the major differences, which deteriorates the performance of a speech utterance verification system severely, is the lengthening of vowels in singing. To solve this problem, this work proposes two improved methods, namely, vowel shrinking and vowel decimation. Both of the methods aim to adjust the length of a vowel in singing to a normal length in speaking. Our experiment show that the proposed method can improve the performance of the previous sung lyrics verification system around 2%-3% in terms of error rate reduction.