An Effective Approach for Vocal Melody Extraction from Polyphonic Music on GPU

Guangchao Yao (1, 2) , Yao Zheng (1, 2) , Limin Xiao (1) , Li Ruan (1, 2) , Zhen Lin (1, 2) , Junjie Peng (3)

1 State key Laboratory of Software Development Environment [Beijing]
2 School of Computer Science and Engineering [Beijing]
3 School of Computer Engineering and Science [Shanghai]

Melody extraction from polyphonic music is a valuable but difficult problem in music information retrieval. The extraction incurs a large computational cost that limits its application. Growing processing cores and increased bandwidth have made GPU an ideal candidate for the development of fine-grained parallel algorithms. In this paper, we present a parallel approach for salience-based melody extraction from polyphonic music using CUDA. For 21 seconds of polyphonic clip, the extraction time is cut from 3 seconds to 33 milliseconds using NVIDIA GeForce GTX 480 which is up to 100 times faster. The increased performance allows the melody extraction to be carried out for real-time applications. Furthermore, the evaluation of the extraction on huge datasets is also possible. We give insight into how such significant speed gains are made and encourage the development and adoption of GPU in music information retrieval field.

Domaines: Informatique [cs]

Guangchao Yao, Yao Zheng, Limin Xiao, Li Ruan, Zhen Lin, et al.. An Effective Approach for Vocal Melody Extraction from Polyphonic Music on GPU. 10th International Conference on Network and Parallel Computing (NPC), Sep 2013, Guiyang, China. pp.284-297, ⟨10.1007/978-3-642-40820-5_24⟩. ⟨hal-01513779⟩