How much improvement in training quality can I expect from using a 3-hour audio dataset compared to a 30-minute one?

[Nisekoi-1]

Is the improvement in training quality worth the extra time required?

[Chi8wah]

Due to this video, it's said that dataset longer than 30 minutes will begin to show boundary effects, but based on my own practice, I personally recommend the longer the better, even though it may be a bit difficult to distinguish the difference between the two.