歌詞演唱錯誤偵測

Automatic Sung Lyrics Verification

本研究嘗試發展一種唱詞確認系統，以自動判斷演唱者是否唱錯歌詞。雖然直覺上，唱詞確認相似於語句確認問題，可以利用語音辨認上所使用的方法來處理，但由於歌唱聲音訊號就像語音訊號的伸縮、變形過後版本，我們發現直接利用語句確認進行唱詞確認的效果並不如預期。有鑑於歌唱時常因母音被拉長若干倍而造成與說話時的訊號相差甚多，我們試圖找出歌唱中的母音位置，並對其長度壓縮或裁剪，使其接近語音訊號，以使語句確認方法較能正常運作。經實驗結果顯示，透過母音長度壓縮或裁剪可大幅提升唱詞判斷的正確率。

This study proposes a sung lyrics verification system for detecting if the lyrics sung by a performer are incorrect and further pointing out the potential mistake that the performer made. In essence, sung lyrics verification is similar to the problem of speech utterance verification in the speech recognition research community, and therefore the techniques in the letter can be applied to the former. However, our preliminary experiment found that a speech utterance verification system cannot handle singing data well, mainly because of the significant differences between singing and speech. To tackle this problem, we develop two strategies, respectively, from a signal processing perspective and from a model processing perspective. In the signal processing, recognizing that the vowels are often lengthened during singing, we propose vowel shrinking and vowel decimation to adjust the length of a vowel in singing to a normal length in speaking. In the model processing, we include a duration model concept in the acoustic modeling to reduce the differences between singing and speech. Our experiments show that the proposed methods can improve the performance of the sung lyrics verification to 72% and 90% accuracy using vowel shrinking, vowel decimation, and duration model approach, respectively, compared to 63% accuracy obtained with the baseline speech utterance verification system.