songronglee / mir-svc Goto Github PK

I use python inference.py $src_file $target_dir $singer_id $model_path --pitch-shift $pitch_shift --two-phase --train-data-dir $train_data_dir to inference ADIZ 13.WAV convert singer JTAN. It neet 15hour on 2080Ti? Maybe is ture?

Hi, how does the speaker classification loss look like on your side? It diverges in the middle of the training process in my experiment. Any idea about how to balance the speaker classification loss for the classification network(positive) and the main network(negative)?

See the training loss curve for speaker classification below. (My network follows the ideas of speaker classification and pitch regression modules, but not the exact one as your task.)

Hello there. I'm trying to train on my own audio, however when I train I got

Model read from proposed.
Convertor initialized.
Creating datasets...
Training data dataset initialized from data
Training...
Training done in 0.0 minutes.

I did the following steps:

1. Created a dir called raw_dir and added folders for my singer (just one) and then the audios in it.
2. I ran python data_augmentation.py raw_dir aug_dir
3. I ran python data_preprocess.py aug_dir data
4. Try to train with python train.py data checkpoints --model-path ./proposed

This is how my raw_dir looks like:

-> $raw_dir/
├── singer
│   ├── 01.wav

My idea is to have this new singer available to use as a target to convert other audio to its style.