Using custom data for model training and testing

[leandermaben]

I need to train and evaluate the ConvTasNet model on speech enhancement for a custom dataset. How should I go about formatting the data, and what changes need to be made in the run.sh file.

[mpariente]

Hey, thanks for the issue !

Have you had a look at already implemented speech enhancement dataset, how do they look like ?

[leandermaben]

Hi, thanks for the response. By dataset, I don't mean the dataset class. I mean I want to use my own data ( clean and noisy .wav files) for training instead of wham or any of the other ones. I was hoping to do this directly using the ConvTasNet recipe by just changing the parameters in run.sh. Would this be possible?

[mpariente]

Not directly, not. Unless you create the json file that the WHAM dataset expects for your data, which is quite simple.

Otherwise, you need to create a dataset class for your own data, that's why I suggest you to look at how the datasets are implemented.

[leandermaben]

Sure, I will look into it. Thank You.