Singing Voice Detection Algorithm Based on a Squeeze-and-Excitation Residual Network

In this paper, we proposed an algorithm based on the squeeze-and-excitation residual network. Other than those algorithms armed with complicated feature engineering, the proposed network could learn more effective features by the hierarchical convolution collaborated with the squeeze-and-excitation operation, only fed with the naive acoustic feature. In this algorithm, the residual structure can easily extend the depth of convolutional network, and the squeeze-and-excitation operation can fuse the learned multiple features by the adjusted weights, and furtherly can improve the overall performances. To prove the feasibility and effectiveness, we conducted the experiments on the two public datasets. Compared with one of the state of the art base line, the proposed algorithm produced the significantly better performance.

In Journal of Fudan University (Natural Science)