Lecture Notes in Electrical Engineering
Using audio fingerprint for representing the audio content is a simple way for detecting similar songs with increasing the accuracy. This becomes more important problem when we need to handle ten million songs in the Internet. Not only the goodness of audio fingerprint extraction algorithm, but also the searching algorithm can affect much the effectiveness of searching system. In previous work, we proposed a new massively parallel system that can handle the audio fingerprint searching problem for thousands of queries at the same time based on HiFP2.0 audio fingerprint extraction algorithm. Our system uses LSH and K-modes for combining the data-flow from CPU to GPGPUs. In this paper, we continue proposing methods for increasing the accuracy of our system and increasing efficiency of massively parallel. We also propose new cluster algorithm extended from K-modes that can meet the requirements for GPGPU system with different size devices.
