Learning to Align Polyphonic Music

We describe an efficient learning algorithm for aligning a symbolic representation of a musical piece with its acoustic counterpart. Our method employs a supervised learning approach by using a training set of aligned symbolic and acoustic representations. The alignment function we devise is based on mapping the input acousticsymbolic representation along with the target alignment into an abstract vector-space. Building on techniques used for learning support vector machines (SVM), our alignment function distills to a classifier in the abstract vectorspace which separates correct alignments from incorrect ones. We describe a simple iterative algorithm for learning the alignment function and discuss its formal properties. We use our method for aligning MIDI and MP3 representations of polyphonic recordings of piano music. We also compare our discriminative approach to a generative method based on a generalization of hidden Markov models. In all of our experiments, the discriminative method outperforms the HMM-based method.