Automatic Detection of Mispronounced Lyrics in Singing

In this study, we propose an automatic system for detecting mispronounced lyrics in singing, thereby providing information for singing performance assessment. The system is built upon the basis of speech utterance verification and further improved by considering the difference between singing and speech. We recognize that the vowels are often lengthened during singing and thus include a duration modeling concept in the acoustic modeling to absorb the variation of the length of a vowel in singing. Our experiments show that the proposed methods can achieve 11.3% equal error rate in detecting the mispronounced lyrics in singing.