Loss does not converge on verification set

[timobolkart]

Loss does not converge on verification set

[youngstu]

Are there some special tricks used for model training?

[youngstu]

Could adding data improve the generalization ability of the model?
The training data of voca is only 40 sentences per person, which can not be enough.

[youngstu]

Does audio data have requirements for speech speed, volume, pitch and timbre?

[TimoBolkart]

We share the training data and the whole training code, all "tricks" are part of the code. The general problem of the error not reducing on the validation set is the big difference in speaking style between subjects. I recommend using the validation set to qualitatively assess the training progress, the quantitative evaluation is less meaningful.