We consider in this paper sparse audio coding as an alternative to transform audio coding for efficient MIR in the transform domain. We use an existing audio coder based on a sparse representation in a union of MDCT bases, and propose a fast algorithm to compute mid-level representations for beat tracking and chord recognition, respectively an onset detection function and a chromagram. The resulting transform domain system is significantly faster than a comparable state-of-the-art system while obtaining close performance above 8 kbps.