[ssl/bestrq] happy ending---stable training

[Mddct]

No description provided.

[ncakhoa]

@Mddct Can you explain when you stack fbank feature, why you change from normalizing freq dimension to time dimension and please provide some training results for comparing two approaches. I think normalizing freq dimension is more correct than time dimension, since the model wants to learn local distributions.

[Mddct]

@Mddct Can you explain when you stack fbank feature, why you change from normalizing freq dimension to time dimension and please provide some training results for comparing two approaches. I think normalizing freq dimension is more correct than time dimension, since the model wants to learn local distributions.

Thank you for your attention

The normalization of input is just to get richer codes id. The stack input dim=T is normalized here to simulate the calculation method of cmvn of stack input, and experiments show that this normalization will be better than norm. d has a larger proportion of unique codes

[ncakhoa]

Can you show proportion of unique codes in two approaches. Also, have you tried to downstream model.

[Mddct]

Can you show proportion of unique codes in two approaches. Also, have you tried to downstream model.

Yes, some promising results will be released recently.