ValueError: n_wins 2072 > max_length 1300

[LXP-Never]

Do I need to retrain the model with a larger max_length?

[gabrielmittag]

Hi,

No, you could just overwrite the maximum length in the config file. That being said, the model is not really made to estimate such long speech samples.

You can overwrite the config of the checkpoint and write a new checkpoint file with the model weights as follows:

import torch
checkpoint = torch.load('weights/nisqa.tar')
checkpoint['args']['ms_max_segments'] = 3000
torch.save(checkpoint, 'weights/nisqa_2.tar')

Then you can run the model with the new checkpoint:

python run_predict.py --mode predict_file --pretrained_model weights/nisqa_2.tar --deg /path/to/wav/file.wav 

Let me know if it works!

[LXP-Never]

it works! Thank you

[gabrielmittag]

Great!