Signal + Context = Better Classification

Typical signal-based approaches to extract musical descriptions from audio signals only have limited precision. A possible explanation is that they do not exploit take any account of context, which provide important cues in human cognitive processing of music: e.g. electric guitar is unlikely in 1930s music, children choirs rarely perform heavy metal, etc. We propose an architecture to train a large set of binary classifiers simultaneously, for many different musical metadata (genre, instrument, mood, etc.), in such a way that correlations between metadata are used to reinforce each individual classifier. The system is iterative: it uses classification decisions it made on some metadata problems as new features for new, harder classification problems; and hybrid: it uses a signal classifier based on timbre similarity to bootstrap symbolic reasoning with decision trees. While further work is needed, the approach seems to outperform signalonly algorithms by 5% precision on average, and sometimes up to 15% for traditionally difficult problems such as cultural and subjective categories.