Combining pitch-based inference and non-negative spectrogram factorization in separating vocals from polyphonic music.