search

JusperLee / CTCNet

An Audio-Visual Speech Separation Model Inspired by Cortico-Thalamo-Cortical Circuits
Apache License 2.0
70 stars 17 forks source link

the trainning's loss #16

Closed xlzhou01 closed 1 week ago

xlzhou01 commented 1 week ago

我想问一下，您论文中使用的SI-SNR损失就是你train_ctc.y中的pairwise_neg_snr？它是要比pairwise_neg_sisdr训练效果要好一些吗？求解

JusperLee commented 3 days ago

一般来说，snr比sisnr要更好，因为sisnr对于能量不够敏感。会影响输出的能量大小