Speeding up audio fingerprinting over GPUs

This paper presents the use of GPUs (graphic processing units) for implementing an efficient audio fingerprinting (AFP) system for audio music retrieval. Such a music retrieval system can compare a 10-second recording of exact but noisy audio clip to the database of more than 100K songs on a single PC with GPU cards. Due to the use of GPUs, we can achieve a speedup factor of 14 for audio comparison, leading to a satisfactory response time for a practical AFP service. In order to squeeze the computing power out of GPU, we have tried detailed performance analysis to find the best memory usage strategy and the corresponding optimum GPU parameters. As far as we know, this is the first technical paper covering the detailed use and analysis of GPU for audio fingerprinting over a database of more than 100K songs. The resulting system, now with around 203K songs, has been successful deployed on the web for public access.