Hmm, the validity of this claim really depends on what the diff sounds like. Different production batches would give slightly different filter corner frequencies, or different asymmetry in saturation, or different program-dependent ballistics. If the model adds -60dB of quantisation noise, or a whistling signal, or extra band-limiting... that's not equally likely to go unnoticed!Ultimately, APNN achieves an error signal of phase cancellation at about -40dB to -75dB (on the training set, not on the validation set). With this exceptional level of error control, which surpasses even the variance between different production batches of the same hardware model, we can confidently say that APNN is capable of 'deceiving' human ears.
But if you have got the "artifact-like" errors to a lower level then this sounds very promising! Will be interesting to trial.
Statistics: Posted by imrae — Thu Jun 06, 2024 7:53 am