This work addresses the recurring challenge of real-time monophonic and polyphonic audio source classification. The whole normalized power spectrum (NPS) is directly involved in the proposed process, avoiding complex and hazardous traditional feature extraction. It is also a natural candidate for polyphonic events thanks to its additive property in such cases. The classification task is performed through a nonparametric kernel-based generative modeling of the power spectrum. Advantage of this model is twofold: it is almost hypothesis free and it allows to straightforwardly obtain the maximum a posteriori classification rule of online signals. Moreover it makes use of the monophonic dataset to build the polyphonic one. Then, to reach the real-time target, the complexity of the method can be tuned by using a standard hierarchical clustering preprocessing of the prototypes, revealing a particularly efficient computation time and classification accuracy trade-off. The proposed method, called RARE (for Real-time Audio Recognition Engine) reveals encouraging results both in monophonic and polyphonic classification tasks on benchmark and owned datasets, including also the targeted real-time situation. In particular, this method benefits from several advantages compared to the state-of-the-art methods including a reduced training time, no feature extraction, the ability to control the computation - accuracy trade-off and no training on already mixed sounds for polyphonic classification.