After training this model on 175 samples as the training set in the original paper for only ~4500 steps, I am able to achieve ~22 GNSDR for the vocals, which is considerably higher than reported in the paper. I am wondering if there is some error in bss_evaluation that is causing this abnormally high score. Has anyone else reported on the results of this model?
Hi,
After training this model on 175 samples as the training set in the original paper for only ~4500 steps, I am able to achieve ~22 GNSDR for the vocals, which is considerably higher than reported in the paper. I am wondering if there is some error in bss_evaluation that is causing this abnormally high score. Has anyone else reported on the results of this model?
Thanks!