Systems and methods are provided herein relating to audio matching. Interest points that are onsets are generally very efficient in audio matching in that they are robust to multiple types of distortion. Prominent onsets can be detected within an audio signal excerpt as interest points and combined as a function of a set of interest points to form a descriptor. Descriptors associated with an audio signal excerpt that contain a set of prominent onsets as interest points can be used in matching the audio signal excerpt to an audio reference. The benefits in generating and using prominent onsets within descriptors improve the accuracy of an audio matching system.