Intelligibility of Sung Lyrics: A Pilot Study

We propose a system to automatically assess the intelligibility of sung lyrics. We are particularly interested in being able to identify songs which are intelligible to second language learners, as such individuals often sing along the song to help them learn their second language, but this is only helpful if the song is intelligible enough for them to understand. As no automatic system for identifying ‘intelligible’ songs currently exists, songs for second language learners are generally selected by hand, a timeconsuming and onerous process. We conducted an experiment in which test subjects, all of whom are learning English as a second language, were presented with 100 excerpts of songs drawn from five different genres. The test subjects listened to and transcribed the excerpts and the intelligibility of each excerpt was assessed based on average transcription accuracy across subjects. Excerpts that were more accurately transcribed on average were considered to be more intelligible than those less accurately transcribed on average. We then tested standard acoustic features to determine which were most strongly correlated with intelligibility. Our final system classifies the intelligibility of the excerpts and achieves 66% accuracy for 3 classes of intelligibility.