Center for Computer Research in Music and Acoustics

We are starting to think about audio and EEG signals as combinations of sources. This has been true of both images and audio for a while, and an important breakthrough in these types of analysis is the realization that the sources must be strictly positive. Algorithms such as PCA and the FFT divvy up the energy in a signal using both positive and negative components. Non-negative matrix factorization (NMF) creates an alternative explanation of the data where all the factors are strictly positive, since objects can only contribute positive energy to the overall signal.

Slim Essid will be at CCRMA to talk about using NMF to analyze audio signals for auditory scene analysis, and to decompose EEG signals into their independent sources. Both important tasks.