Most of speech/music discrimination techniques proposed in the literature need a great amount of training data in order to provide acceptable results. Besides, they are usually context-dependent. In this paper, we propose a novel technique for speech/music discrimination which relies on first order sound spectrum's statistics as feature vector and a neural network for classification. Experiments driven on 20000 seconds of various audio data show that the proposed technique has a great ability of generalization since a classification accuracy of 96% has been achieved only after a training phase on 80 seconds audio data. Furthermore, the proposed technique is context-independent as it can be applied to various audio sources.