Reducing Mode Collapse

[late marlin]

Creating voice models is new to me, so I want to learn more about mode collapses.
How does a mode collapse affect the end result of a model? How many mode collapses should one expect? Some say a few are fine; others mention discarding their models after seeing one. How does batch size play a part in this, and how can one alter one's dataset to prevent this?

[formal ore]

Seems like your model flatlined way too soon, so I believe it is more of a batch size issue rather than a dataset issue.

To check if it's a dataset issue, try training the model on batch 4. If this doesn't solve it, then it is most likely a dataset issue.

If it is a dataset issue, it could be either low quality (though I'm not all too sure about this, only heard one source about it) or badly trimmed silences.

[late marlin]

Thanks for the help. Is it always better to train on low batch sizes? What does a badly trimmed silence look like?

[formal ore]

1. Not necessarily. Every model has its "golden" batch size, which is the one that has the biggest first drop in loss/g/total.

2. Something over 50 ms can already cause issues afaik. You should trim silences to 0,04 or 0,05s on Audacity

[late marlin]

That could be one of my problems. In my datasets, I always use truncate silence and set silence to zero.

Is that a bad decision?

Also, is it better to leave coughs and such in datasets? I usually leave them out.

[lost onyx]

so it doesnt read the whole dataset and gets stuck on the silent parts to avoid it truncate data and noise gate it , it negatively impacts model ability to accurately represent the full distribution of dataset