Improving MIDI Guitar's Accuracy with NMF and Neural Net

In this paper, we propose a method for improving the accuracy of MIDI guitars. MIDI guitars are useful tools for various purposes from inputting MIDI data to enjoying a jam session system, but existing MIDI guitars do not have sufficient accuracy in converting the performance to an MIDI form. In this paper, we make an attempt on improving the accuracy of a MIDI guitar by integrating it with an audio transcription method based on non-negative matrix factorization (NMF). First, we investigate an NMF-based algorithm for transcribing guitar performances. Although the NMF is a promising method, an effective post-process (i.e., converting the NMF’s output to an MIDI form) is a non-trivial problem. We propose use of a neural network for this conversion. Next, we investigate a method for integrating the outputs of the MIDI guitar and NMF. Because they have different tendencies in wrong outputs, we take an policy of outputting only common parts in the two outputs. Experimental results showed that the F-score of our method was 0.626 whereas those of the MIDI-guitar-only and NMF-and-neural-network-only methods were 0.347 and 0.526, respectively.