基于高分辨率网络和自注意力机制的歌声分离算法

Singing voice separation algorithm based on high resolution network and self-attention mechanism

针对现有歌声分离算法分离精度不高的问题,提出了一种基于高分辨率网络和自注意力机制的歌声分离算法。该算法构建了基于频域模型的深度神经网络,将高分辨率网络作为主干网络,以此保证分离精度,并在网络中融入自注意力机制来捕获歌曲中的重复旋律。在歌声分离算法中,首先通过短时傅里叶变换对音乐信号进行时频转换,得到幅值谱;其次通过构建的神经网络将歌曲幅值谱进行分离,得到人声和伴奏的幅值谱;最后结合原歌曲的相位谱,通过短时傅里叶逆变换得到人声和伴奏的时域信号。结果表明:在MUSDB18数据集上,分离得到的人声和伴奏信号偏差比指标分别为7.68 dB和12.85 dB,相比于基准模型分别提高了21.52%和1.26%。该算法可以增强神经网络特征表达能力,有效提升歌声分离效果。

To address the problem of low separation accuracy of the existing singing voice separation algorithms, a singing voice separation algorithm based on high-resolution network and self-attention mechanism was proposed, which constructed a deep neural network based on the frequency-domain model, used high-resolution network as the backbone network to ensure the separation accuracy, and integrated the self-attention mechanism into the network to capture the repeated melody in the song. The process of singing voice separation algorithm is as follows: Firstly, the short-time Fourier transform was used for the time-frequency transformation of music signal to get the amplitude spectrogram;second, the amplitude spectrum of song was separated by the established neural network to obtain the amplitude spectrogram of the singing voice and accompaniment;finally, the time domain signals of singing voice and accompaniment were obtained by short-time inverse Fourier transform according to the phase spectrogram of the original song. The experimental results show that: on the MUSDB18 dataset, the signal-to-deviation ratio index of singing voice and accompaniment is 7.68 db and 12.85 db respectively, an increase of 21.52% and 1.26% than the benchmark model, indicating that the algorithm proposed in this study can strengthen the feature expression ability of neural network, and effectively improve the effect of singing voice separation.