This thesis investigates new signal representations for audio coding. Existing state-of-the-art audio coders are based either on a transform (transform coding), or on a parametric model (parametric coding), or on a combination of both (hybrid coding). On the one hand, transform coding achieves (near-)transparent quality at high bitrates (e.g. AAC at 64kbps/channel), but gives poor performance at lower bitrates. On the other hand, parametric and hybrid coding achieve better performance than transform coding at low bitrates but cannot give transparent quality at high bitrates. The new approach for signal representation that we propose allows to achieve transparent quality at high bitrates, while giving better performance than transform coding at low bitrates. This signal representation is based on an overcomplete set of time-frequency functions composed by a union of several MDCT bases with different scales. The first major contribution of this thesis is a fast and efficient algorithm that decomposes a signal into this overcomplete set of functions. The second major contribution of this thesis is a set of techniques that allows the coding of these representations in an efficient and scalable way. Finally, this thesis investigates the application to audio indexing. We show that using a union of several MDCT bases allows to go beyond the limitations of the representations used in the transform coders (particularly the frequency resolution), which makes possible an efficient indexing in the transform domain.
