This paper investigated a new scheme for single-sensor audio source separation. This framework is introduced comparatively to the existing Gaussian mixture model generative approach and is focusing on the mixture states rather than on the source states, resulting in a discrete, joint state discriminant approach. The study establishes the theoretical performance bounds of the proposed scheme and an actual source separation system is designed. The performance is computed on a set of musical recordings and a discussion is proposed, including the question of the source correlation and the possible drawbacks of the method.