A increasing amount of studies have recently focused on the use of sparse models for the representation of audio content, as thay allow a significant dimensionality reduction - a desirable quality for machine learning algorithms applied to music analysis tasks. In this talk we focus on the use of sparse coding methods for actual audio coding, and we discuss the choices of model / dictionaries that are the most relevant for a target bitrate / quality. In practice, we have constructed an audio codec with a dictionary made by concatenation of 8-scales MDCT, that outperforms state-of-the-art codecs at very low bitrates for ... sparse sounds (and with similar results on other sounds). An additional feature of the obtained compressed decompositions is that partial decoding can also be used for some complexity-scalable Music Information Retrieval tasks, with some strikingly good results at extremely low bitrates.
This is joint work with E. Ravelli and G. Richard.