My suggestion is to compare the critics rating for the real data and the generated data. Compute the mean of the ratings on real data to get some kind of threshold. If your critic is designed so that a high rating indicates a real sample, then generated data with ratings greater than the mean of the ratings of the real data could be viewed as good enough to ...
Sorry cannot directly reply to your comment as I posted without an account, and you were right! I replaced transposed layers with Upscale1D+Conv1D and that solved the issue.
gen = Conv1DTranspose(128, 4, strides=2, padding='same', kernel_initializer=w_init, use_bias=None)(gen)
should become (notice that strides=2 becomes strides=1):
gen = Upscale1D()(gen)
Because it is possible to fool many different models at once.
See table 2 in this paper, for an example using adversarial perturbations: https://arxiv.org/pdf/1610.08401.pdf
That being said, there is no reason to think that using two detectors at once will not increase chance to detect deepfakes. It will just not resolve the problem completely.