Note Recognition from Monophonic Audio: A Clustering Approach

摘要

We describe a new method for recognizing notes from mono-phonic audio, such as sung or whistled queries. Our method achieves results similar to known methods, but without any probabilistic models that would need to be trained. Instead, we define a distance function for audio frames that captures three criteria of closeness which usually coincide with frames belonging to the same note: small pitch difference, small loudness fluctuations between the frames, and the absence of non-pitched frames between the compared frames. We use this distance function for clustering frames such that the total intra-cluster costs are minimized. Criteria for clustering termination include the uniformity of note costs. This new method is fast, does not rely on any particular fundamental frequency estimation method being used, and it is largely independent of the input mode (singing, whistling, playing an instrument). It is already being used successfully for the "query by humming/whistling/playing" search feature on the publicly available collaborative melody directory Musipedia.org.