User-Specific Learning for Recognizing a Singer’s Intended Pitch

We consider the problem of automatic vocal melody transcription: translating an audio recording of a sung melody into a musical score. While previous work has focused on finding the closest notes to the singer’s tracked pitch, we instead seek to recover the melody the singer intended to sing. Often, the melody a singer intended to sing differs from what they actually sang; our hypothesis is that this occurs in a singer-specific way. For example, a given singer may often be flat in certain parts of her range, or another may have difficulty with certain intervals. We thus pursue methods for singer-specific training which use learning to combine different methods for pitch prediction. In our experiments with human subjects, we show that via a short training procedure we can learn a singer-specific pitch predictor and significantly improve transcription of intended pitch over other methods. For an average user, our method gives a 20 to 30 percent reduction in pitch classification errors with respect to a baseline method which is comparable to commercial voice transcription tools. For some users, we achieve even more dramatic reductions. Our best results come from a combination of singer-specific-learning with non-singer-specific feature selection. We also discuss the implications of our work for training more general control signals. We make our experimental data available to allow others to replicate or extend our results.