search

tsurumeso / vocal-remover

Vocal Remover using Deep Neural Networks
MIT License
1.55k stars 222 forks source link

Quality Reduction in v3 #55

Closed Anjok07 closed 4 years ago

Anjok07 commented 4 years ago

Hello,

Quick question, I noticed there seems to be a greater reduction in the quality of the accompaniment compared to v2. Was this to make the vocal removal more aggressive? If so, what parameters need to be tweaked to bring the accompaniment quality back up to v2?

Thanks!

404000 commented 4 years ago

Anjok07 in version 3.0.1 tsurumuso wrote than if you want to improve the vocal removal quality with TTA, use the --tta option Does this tta improve accompaniment quality?

Anjok07 commented 4 years ago

Anjok07 in version 3.0.1 tsurumuso wrote than if you want to improve the vocal removal quality with TTA, use the --tta option Does this tta improve accompaniment quality?

Not really. The quality of the accompaniment isn't even close to v2, even with the tta option on.

404000 commented 4 years ago

How did you know that the quality of accompaniment decreased so much?

tsurumeso commented 4 years ago

Was this to make the vocal removal more aggressive? If so, what parameters need to be tweaked to bring the accompaniment quality back up to v2?

Probably so. You can adjust the vocal removal level with the --max_reduction_rate option. By setting this to 0, the behavior is almost the same as v2.

404000 commented 4 years ago

Tsurumeso Will it be like this?python inference.py --input path/to/an/audio/file --max_reduction_rate 0

tsurumeso commented 4 years ago

It's a training option. It will be like this:

python train.py --dataset path/to/dataset --max_reduction_rate 0 --mixup_rate 0.5 --gpu 0
Anjok07 commented 4 years ago

Thank you for the feedback! I will try this

tsurumeso commented 4 years ago

Sorry, there was a bug in v3.0.3. Please use v3.0.4 as follows:

python train.py --dataset path/to/dataset --reduction_rate 0 --mixup_rate 0.5 --gpu 0