Using Oracle Analysis for Decomposition-Based Automatic Music Transcription

Abstract

One approach to Automatic Music Transcription (AMT) is to decompose a spectrogram with a dictionary matrix that contains a pitch-labelled note spectrum atom in each column. AMT performance is typically measured using frame-based comparison, while an alternative perspective is to use an event-based analysis. We have previously proposed an AMT system, based on the use of structured sparse representations. The method is described and experimental results are given, which are seen to be promising. An inspection of the graphical AMT output known as a piano roll may lead one to think that the performance may be slightly better than is suggested by the AMT metrics used. This leads us to perform an oracle analysis of the AMT system, with some interesting outcomes which may have implications for decomposition based AMT in general

Non-Negative Group Sparsity with Subspace Note Modelling for Polyphonic Transcription

“…Alternatively, the results may be interpreted in terms of the quality of model. The subspace dictionary provides a more descriptive model than the AHD by including implicit temporal information, while the approximate additivity assumed by NMF-based AMT may be less effective in the presence of transients and onsets [57]. From this perspective, GBF-NNLS performs well when the spectra can be well-approximated, as also seen in Table I, but is not so robust as the NMF-based methods which outperfom GBF-NNLS in other cases.…”

“…Structured sparse decomposition methods may provide further possibilities. It is easy to observe in NMF-based AMT the problem of low-energy elements of sustained notes being overpowered by false positives related to higher energy active notes [57]. We previously observed improved AMT, in both analyses, simply by using a low offset threshold, clustering adjacent active atoms into molecules, and subsequently determining activation over a whole molecule rather than for individual atoms [15] [37], using stepwise methods.…”

“…We have previously observed systematic problems with the described onset detector [57], including not capturing retriggered notes and detecting spurious false onsets when the activation level is near the threshold. We propose some modifications.…”