unilight/ldnet

Official implementation of the paper: "LDNet: Unified Listener Dependent Modeling in MOS Prediction for Synthetic Speech"

ldnet's Issues

How to get the labels of BVCC test wav?

How can I get BVCC test dataset? I only found data and label of train and val in VoiceMOS codalab, but I don't know how to get the labels of test wav. Is it opensource?

Expected 4 got 3

[Info] Use mean listener: False
[Info] Model parameters: 1386198
Train: 100% 845/845 [06:23<00:00, 2.20 step/s]tep/s, loss=4.78, mean_loss=1.32, LD_loss=0.866]
Valid: 0% 0/3000 [00:01<?, ? step/s], 2.46 step/s, loss=6.95, mean_loss=1.48, LD_loss=1.37]
Train: 18% 154/845 [01:26<06:29, 1.77 step/s]
Traceback (most recent call last):p/s]
File "train.py", line 269, in
main()
File "train.py", line 265, in main
valid(config["inference_mode"], model, valid_loader, valid_set.systems, save_dir, global_step, "Valid")
File "train.py", line 36, in valid
mag_sgrams_padded, avg_scores, sys_names, wav_names = batch
ValueError: not enough values to unpack (expected 4, got 3)
Overall: 2% 999/50000 [07:50<6:24:29, 2.12 step/s, loss=6.95, mean_loss=1.48, LD_loss=1.37]

While running python train.py --config configs/MBNet.yaml --tag model1

vcc2018_training_data.csv is not 13580

Hello, I have found that there are only 54040/4=13510 data entries in the vcc2018_training_data.csv file that you have publicly released. However, the paper states that there are 13580 data entries. Do you have the complete 13580 file? Or do you know where I should get it from?