Monophony vs Polyphony: A New Method Based on Weibull Bivariate Models

Our contribution takes place in the context of music indexation. In many applications, such as multipitch estimation, it can be useful to know the number of notes played at a time. In this work, we aim at distinguish monophonies (one note at a time) from polyphonies (several notes at a time). We analyze an indicator which gives the confidence on the estimated pitch. In the case of a monophony, the pitch is relatively easy to determine, this indicator is low. In the case of a polyphony, the pitch is much more difficult to determine, so the indicator is higher and varies more. Considering these two facts, we compute the short term mean and variance of the indicator, and model the bivariate repartition of these two parameters with Weibull bivariate distributions for each class (monophony and polyphony). The classification is made by computing the likelihood over one second for each class and taking the best one. Models are learned with 25 seconds of each kind of signal. Our best results give a global error rate of 6.3 %, performed on a balanced corpus containing approximately 18 minutes of signal.