An on-line algorithm of guitar performance transcription using non-negative matrix factorization

We describe an on-line algorithm for transcribing guitar performances using non-negative matrix factorization (NMF). NMF is a promising method for polyphonic music transcription, but is originally unsuitable for an on-line algorithm because it requires the whole (from beginning to end) waveform. In this paper, we propose a method using two different preliminary performances. First, the basis vector (i.e., spectrum) for each note is estimated from the first preliminary performance. Second, the gain matrix is calculated with the (supervised) NMF using the estimated basis vectors. Finally, a MIDI sequence is generated by thresholding the gain matrix, where the threshold had been learned in advance using the second preliminary performance. Experimental results show that our method improves the accuracy of guitar performance transcription.