Section: New Results
Real-time Polyphonic Music Recognition
We investigated real-time recognition of overlapping music events in two context of dictionary-based detection and real-time alignment:
Real-time detection of overlapping sound events using non-negative matrix factorization
Participants : Arnaud Dessein, Arshia Cont.
Non-negative matrix factorization (NMF) methods have naturally found their way since their inception to sound and music processing. This work is an extension to our previous work in  on Real-time Music Transcription using sparse NMF methods. We investigate the problem of real-time detection of overlapping sound events by employing NMF techniques. We consider a setup where audio streams arrive in real-time to the system and are decomposed onto a dictionary of event templates learned off-line prior to the decomposition. An important drawback of existing approaches in this context is the lack of controls on the decomposition. We propose and compare two provably convergent algorithms that address this issue, by controlling respectively the sparsity of the decomposition and the trade-off of the decomposition between the different frequency components. Sparsity regularization is considered in the framework of convex quadratic programming, while frequency compromise is introduced by employing the beta-divergence as a cost function. The two algorithms are evaluated on the multi-source detection tasks of polyphonic music transcription, drum transcription and environmental sound recognition. The obtained results in  show how the proposed approaches can improve detection in such applications, while maintaining low computational costs that are suitable for real-time.
A specialized version of NMF for Real-time Music Transcription is exposed in Arnaud Dessein's PhD thesis  .
These methods will be subject to software development in 2013.
Robust Real-time Polyphonic Audio-to-Score Alignment
Participant : Arshia Cont.
The Antescofo system is polyphonic since 2009 but its use in highly polyphonic and noisy concert environments have been challenging. To overcome this, we have studied more robust inference mechanisms. As a results, the previous inference mechanism based on maximum a posteriori of Viterbi Forward variables in mixed semi-Markov and Markov chains in  were abandoned in favor of a more robust method based on importance resampling on state-space models and smoothing of variable-order hybrid chains. This has led to robust real-time alignment and the employment of the system in various Piano performances in 2012. Further extensions are currently under study.