We present an investigation into signal processing models appropriate for audio, and especially high quality musical signals, by means of Bayesian atomic decompositions. At present, many models rely on short-term stationarity of the audio, or highly limiting forms of non-stationarity. Moreover, they are well-suited only to low-level inference tasks. We seek to formulate a new generation of audio models that will address the main limitations of the existing ones and permit high-level inference. As we show, such models result from the marriage of an overcomplete dictionary of time-frequency atoms with structured hierarchical prior probability distributions developed specifically for audio signals, in order to model coefficient correlation in time and frequency.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.