Syncnet training? Chilcken and egg?

[sylyt62]

I'm a newbie in this field. If my question is stupid, don't laugh at me pls:)

Since syncnet does not perform well on other languages, we need to train it.

To train it, we need to prepare synchronized data.

To prepare synchronized data, we need to filter our data using syncnet.

Qs: How can I get a working syncnet before train it?

ref: https://github.com/primepake/better_wav2lip/issues/21

[ghost]

Hey, you made me remember the first time I'm getting in this field :) Btw, I need to use this repo to do that: https://github.com/joonson/syncnet_python

[sylyt62]

I tried this repo. I used the official syncnet_v2.model, but the performance is really bad on my mandarin dataset. That leads to my question above lol...

[ghost]

it can estimate your data by running this repo can filter noises and misalignment data, of course because of estimate so it's not perfect. So you need to find the range of score that would be fit on your dataset

[conkeur]

I'm a newbie in this field. If my question is stupid, don't laugh at me pls:)

Since syncnet does not perform well on other languages, we need to train it.

To train it, we need to prepare synchronized data.

To prepare synchronized data, we need to filter our data using syncnet.

Qs: How can I get a working syncnet before train it?

ref: #21

I suggest you read the original paper of syncnet,you can train a initial version syncnet with the data containing unsynchronized video.then use this model to filter out the unsynchronized video to retrain a new model.