Automatic music transcription using Non-negative Matrix Factorization

This paper proposes an effective method for the automatic music transcription in polyphonic music. The method consists of a combination of Non-negative Matrix Factorization (NMF), subharmonic summation method and onset detection algorithm. We decompose the magnitude spectrum of a music signal into the spectral component and the temporal information of every note using NMF. Then, the accurate pitch of each note is calculated from the decomposed frequency components based on the subharmonic summation method. And an algorithm for detecting the onset is applied for estimating the temporal information of a musical note. Our method is simple and has a low computational cost, because the method is not a note training-based. The previous researches using NMF detect the pitch and the time duration 'manually', therefore the previous methods are difficult to use in the real engineering. Our proposed method improved this problem with 'automatically' detecting the fundamental frequency and the rhythm component. Furthermore, the proposed method automatically performed the indexing of the musical notes which is useful in the real engineering field. The transcription performance is evaluated with recorded polyphonic music signals, and the performance of the proposed method is better than the conventional NMF based methods in estimating both frequency component and time duration information.