How many samples must a dataset at least contain ?

[Ca-ressemble-a-du-fake]

Hi,
I rephrase and separate my questions (all about dataset samples) now that I have more experience.

A) When training a VITS model how many samples should the dataset at least contain to give decent results ?
A1) Can the dataset have less samples when transfer leaning (ie with restore ) or it does not depend on the method ?

A2) Does VITS need less samples than other models (ex Tacotron2) ?

A3) Should the dataset contain a mix of samples with different length (to be able to synthesize long and short sentences) ?

B) When doing voice conversion how many samples should there be :

B1) In the driving samples ?

B2) In the target samples ?

B3) It seems that only few samples are sufficient (1 minutes). Is it because the samples are only used to compute the embeddings ?

Thanks

[Ca-ressemble-a-du-fake]

While reading other posts, the answer for A) is at least 100 samples although the paper states that they used 15 samples (61 seconds overall).