Fast GPU Audio Identification

Abstract. Audio identification consist in the ability to pair audio sig-nals of the same perceptual nature. In other words, the aim is to be able to compare an audio signal with a modified versions perceptually equivalent. To accomplish that, an audio fingerprint is extracted from the signals and only the fingerprints are compared to asses the simi-larity. Some guarantee have to be given about the equivalence between comparing audio fingerprints and perceptually comparing the signals. In designing AFPs, a dense representation is more robust than a sparse one. A dense representation also imply more compute cycles and hence a slower processing speed. To speedup the computing of a very dense audio fingerprint, able to stand stable under noise, re-recording, low-pass filtering, etc., we propose the use of a massive parallel architecture based on the Graphics Processing Unit (GPU) with the CUDA programming kit. We prove experimentally that even with a relatively small GPU and using a single core in the GPU, we are able to obtain a notable speedup per core in a GPU/CPU model. We compared our FFT implementation against state of the art CUFFT obtaining impressive results, hence our FFT implementation can help other areas of application.